# Alexey Evdokimov: AI product producer **Homepage:** https://evdokimov.ai/en **Telegram channel:** https://t.me/ltunevibe (2,700+ subscribers) **Last updated:** 2026-11-16 --- ## Who I produce AI products. I take a business problem and turn it into a service people use every day. From mapping the processes to launch and support. Alexey Evdokimov, AI architect and producer. I build products end to end: technology, design, launch. Own apps, client systems, hardware. Taught at HSE University. ## What I do ### Find where AI will save you time A list of tasks you can hand over to AI, and the first one we start with. ### Plug a ready-made agent into your process An agent that works inside your process: it tracks deadlines, runs projects and writes texts in your voice. ### Build your own AI service for your task A service built around your process, from prototype to launch, and it stays yours. ### Get a stuck project to launch A straight answer on whether to fix or rebuild, and a version people use every day. ### Teach you and your team to work with AI on your own Two formats. Mentoring: a programme built around your request and a personal page with files and templates. Team training: sessions on the team's real tasks and templates people keep working with on their own. ## Also on the site (in Russian) - [AI agents for business](https://evdokimov.ai/ai-agenty/): building and deploying AI agents - [Routine automation with AI](https://evdokimov.ai/ai-avtomatizaciya/): bringing AI into company processes - [MVP with AI](https://evdokimov.ai/mvp-bez-razrabotchikov/): a working product in 2-4 weeks without an in-house dev team - [Documents and contracts with AI](https://evdokimov.ai/ai-dokumenty/): extraction, checks, risks, redlines - [AI for marketing and creative](https://evdokimov.ai/ai-marketing/): on-brand generation, trend analysis - [GEO: getting named in AI answers](https://evdokimov.ai/geo/): so that ChatGPT, Gemini and Alice name your brand - [Vibe-coding mentorship](https://evdokimov.ai/vibecoding/): how to build and launch your own products with AI - [Lab](https://evdokimov.ai/lab/): step-by-step breakdowns of AI techniques from fresh videos and threads, one every morning ## Work ### LTUNE DAILY - My product · 2026 - Case page: https://evdokimov.ai/en/works/ltune-daily/ A personal newspaper for Kindle: every night AI gathers my news, work briefing and archive photos and delivers it by 6:30 **The problem:** Each night AI lays out a newspaper for my Kindle. By 6:30 an issue is on my e-reader: news from my channels, ad-free, a project briefing and archive photos from this day. I grab the Kindle, not the phone. **What was done:** Built the first working version, from layout to sending from the server, with AI agents in one evening. Then tuned it issue by issue, reading each one on the Kindle. **The hard part:** - Filtering before the editor. Calling a big model on every post is expensive. A fast sorter, Jev, sits before the editor: it writes nothing, only answers closed questions and drops a fifth to a third. - A newspaper, not a blog post. Models love "in this issue we will tell you". Strict rules fixed it: the paper never talks about itself, every paragraph has a hard fact, headlines explain the event. - Kindle broke the layout. Send to Kindle turned a vector PDF into a hundred pages of mush. Now pages render at 2160×2880, photos get a newspaper halftone, the issue ships as images in EPUB. **What it does:** - Collectors: Telegram via web preview without API, media, GitHub releases, industry news. - Briefing from my knowledge base: Tasks, calendar, project statuses and a week of work session summaries. - Jev sorter: Is it an ad, is there substance, which section, how important. - Claude as editor: Writes the issue by the style guide, only big news makes the front. - Magazine-grade layout: New Yorker and the old Afisha as references, Cyrillic-ready fonts only. - Photos of the day: Shots from my archive taken on this day in past years. ### Capybara Counts - Kids game · 2026 - Case page: https://evdokimov.ai/en/works/kapibara/ A phone game for second graders: a capybara teaches two-digit column addition with carrying **The problem:** Second grade, two-digit sums, and a parent stuck too. The capybara walks the child through two-digit sums: mentally, in columns and with carrying. The child enjoys playing, and the grown-up sees what works. **What was done:** Built it end to end in two days: level design, the game scene, the character, voice-over and saving without a server. **The hard part:** - The road to the hot spring. Each sum is a step along the path through a garden, a river, bamboo and a volcano. A new level opens after 8 of 10 correct, so difficulty grows with the child. - A crocodile and mini games. A crocodile who stole the tangerines guards every gate, the child beats him with a run of sums. Every five sums there is a mini game: bubbles or crates into a basket. - Hats and no scolding. Stars buy the capybara hats, there are 12 of them. Mistakes are never scolded: the capybara sees where the child slipped and talks it through on the same tangerines. **What it does:** - Seven levels: From number composition to columns with carrying, mixed sums at the end. - Tangerines and crates: A tangerine is a one, a crate is a ten, the carry is visible. - Everything is voiced: Over 500 lines, so it works even if the child still reads slowly. - Three notations: Peterson-style, tally sticks or digits only, switched on the grown-up screen. - Rewards without penalties: Stars, 12 hats, stickers, a daily chest, a streak that never punishes skipping. - Like an app: Goes on the home screen and works without the internet. **Where to see it:** [kapibara.evdokimov.ai (RU)](https://kapibara.evdokimov.ai) ### SOND - Own product · iOS · live since January 2026 - Case page: https://evdokimov.ai/en/works/sond/ A voice recorder for iPhone that hears what was not said out loud **The problem:** Summaries are everywhere. A real breakdown, nowhere. SOND records the conversation and shows what actually happened: where you and them diverged, what they kept circling back to, the question you forgot to ask. **What was done:** Built end to end: product, iOS and Mac, backend, design, website and launch. **The hard part:** - Meaning lives in the dynamics. Doubt never sounds like the word “doubt”. It shows up when someone returns to budget for the third time. So SOND reads the whole conversation, even a two-hour one, not chunks. - A lecture is not a negotiation. SOND recognizes the kind of conversation. In a negotiation it looks for client signals, in a lecture for contradictions between claims, in a coaching session for patterns people return to. - Ask across everything, give nothing away. Search across recordings runs on the iPhone itself. Only the few passages picked for the answer go to the model, and the answer links to the exact second. **What it does:** - Ask all your recordings: “What did we decide on budget?” A quoted answer that jumps to the moment. - No bot on the call: Your phone sits next to the laptop, no notetaker in the participant list. - Any audio: A YouTube link, a Telegram voice memo, any audio file. - PDF protocol: Looks like a protocol, reads like an article. - Recordings stay on device: No accounts, no passwords. iCloud backup if you turn it on. - iPhone and Mac: One subscription for both. **Where to see it:** [App Store · iPhone](https://apps.apple.com/app/id6758042298), [App Store · Mac](https://apps.apple.com/app/id6760844199), [sond.live](https://sond.live/en/), [API for business](https://sond.live/ru/business) ### SVOD - Construction · 2026 · live - Case page: https://evdokimov.ai/en/works/svod/ An AI service for builders that reads a construction contract and tells you whether it can be signed **The problem:** You can paste a contract into ChatGPT. So why SVOD? In a chat, you decide what to ask. In your account, SVOD follows a construction-contract checklist for your side and finds clauses the contract lacks. **What was done:** Built the whole product: the analysis engine, bid comparison, backend, website, payments and launch. Together with a partner. **The hard part:** - One clause, two truths. A 10% retention suits the customer and hurts the contractor. SVOD first asks which side you're on, then reads the contract the way the other side's lawyer would. - The contract is longer than the model's reply. A model can't retype a 300,000-character contract in full. So it returns only the edits, and the server inserts them into your Word file itself. - Negotiations go in circles. The other side sent back their own version? SVOD checks it against your position and drafts counter-comments and a letter to their lawyer. And so on, until it's signed. **What it does:** - Even a scan or a photo: PDF, Word or a photo of a signed contract. OCR is built in. - Risk map: Critical, warning or fine: every clause comes with a quote from the contract. - Ready Word file: Pick the edits, get a contract with every change tracked. - Protocol of disagreements: Built from the edits you marked, with a cover letter for the other side. - Estimate and bid comparison: Brings bids to one scope and one VAT basis, then shows the best deal. - Data stays in Russia: Database in a Russian data center. The file is deleted right after OCR. **Where to see it:** [svod.legal (RU)](https://svod.legal), [Sample contract review (RU)](https://svod.legal/primer), [Sample estimate and bid comparison (RU)](https://svod.legal/primer/smety), [Protocol of disagreements online (RU)](https://svod.legal/protokol-raznoglasiy), [Construction contract risks: a checklist (RU)](https://svod.legal/riski) ### BUREAU - Device · 2026 - Case page: https://evdokimov.ai/en/works/buro/ A desk console that calls you when an AI agent needs a decision and takes the answer by voice **The problem:** A pixel you that waits for agents instead of you. The desk console watches every AI agent session on your Mac and calls when one needs you. Tap or speak, and the answer lands in exactly that chat. **What was done:** Built it end to end: firmware, the Mac server, the sound and the call, a pixel twin of ~110 frames. **The hard part:** - Figure out who needs you. Each session reports its own state to the console. The queue goes: asking, waiting, stuck, working. The card opens by itself only when an agent asks a question or needs permission. - Get the question in a second. A raw terminal command says nothing. Within seconds a fast model writes the gist on the card: "Deploy the site to Hetzner", with the command in small print below. - The right chat, and only your words. Your voice types into that very session's tab, not wherever the cursor is. Buttons send exactly what's written on them: the console can't speak for you. **What it does:** - Desk: Clock, plan limits as a traffic light, and one line: who's waiting. - Queue review: One at a time: decide, next one up, "1/3" in the header. - Session list: Who's asking, who's thinking and how many minutes they've waited. - Day feed: The last six events in words, with times. - Dictation: Speech without an agent's name goes to the cursor on the Mac. - Living twin: Blinks, listens with a hand to its ear, falls asleep after three quiet minutes. **Where to see it:** [Page in "Objects" (RU)](https://evdokimov.ai/objects/#device) ### RADAR - Production · 2026 · live - Case page: https://evdokimov.ai/en/works/radar/ An AI project tracker that reads work chats on its own and keeps the client informed without reports or calls **The problem:** The client never has to ask "so, where are we?" RADAR reads work chats, emails and meetings, pulls out decisions, promises and deadlines, and keeps a live board open to the client. **What was done:** Came up with it and built it end to end: the parsing pipeline, the client board, the owner's panel and the decisions bot. **The hard part:** - The model can't make up a quote. Every event on the board comes with a quote, and code checks that it's a verbatim piece of the original message. No match, no event. - The client sees status, not the back office. Contractor rates, their names and internal arguments never leak out. Six independent layers of defense, from a field whitelist to a check of the live board every 15 minutes. - It only acts alone when it can undo. The system applies internal steps itself when it's confident, with an undo button. Anything the client will see waits for one "yes" from the owner in the morning card. **What it does:** - Chats, email, meetings: Telegram, emails and meeting notes in one pipeline. - A live board for the client: Stages, deadlines and what we're waiting on, no prices, no back office. - Promises and deadlines: Who promised what and by when, with a reminder before the date. - Morning card: The past day's auto-changes with undo, plus decisions only the owner can make. - Change log: What changed on the board, when, and which message caused it. - Any project: Each project is a config file. Nothing project-specific in the code. ### Second brain - Own system · in daily use - Case page: https://evdokimov.ai/en/works/brain/ A personal base of years of chats, email, notes and meetings, searchable in plain words **The problem:** Notes since 2012, chats since 2015, searchable in plain words. Telegram, email, notes, calendar, Drive, meeting recordings and AI sessions flow into one base. It updates every two hours, and profiles of people and projects build themselves. **What was done:** Built it myself: source collectors, a base with search by meaning, relationship graphs, profiles, a morning briefing and a Telegram bot. **The hard part:** - Links instead of folders. People, projects, meetings and emails are linked in a graph. A question about a person walks through their links: deals, recent chats, meetings, decisions. The answer draws on several sources at once. - Raw data stays untouched. Chats, emails and notes go into the base as is, read-only. Everything derived lives in a separate layer: profiles, indexes, a dashboard. A synthesis error never touches the source. - Newer beats older. Every source has a date. If a profile is from spring but everything changed in a chat yesterday, the answer goes with yesterday and says plainly where the base disagrees. **What it does:** - A pipeline every two hours: New chats, emails, notes and meetings land in the base on their own. - Morning briefing: Meetings, urgent tasks and stuck projects, from real data only. - Profiles with no manual work: For people and projects: history, status and open questions. - Telegram bot: Ask by voice or text from anywhere, get a plain-text answer. - Search by meaning: Finds a note even when you no longer remember the exact words. - My own hardware: The base lives on my own server, not in someone else's cloud. ### SUMA - Tools · 2026 - Case page: https://evdokimov.ai/en/works/suma/ A mini app of Telegram channel digests that shows only what stands out **The problem:** A hundred and thirty channels, half repeating each other. SUMA reads your subscriptions and sends one summary by breakfast: only posts that did better than usual for their channel, with no copy-paste and no other people's roundups. **What was done:** Built it for myself end to end: post collection, the ranking algorithm, the bot, the Mini App and a Mac app. **The hard part:** - A big channel isn't an important one. A post's views are compared with its channel's median on a log scale, a z-score. A strong post from a small channel beats a filler post from a million-subscriber one. - Freshness without zeroing out. A post's weight halves in about 18 hours: around 60% after 12 hours, 40% after a day. It never drops below 20%, so a strong post from yesterday doesn't vanish. - No monopoly, no retellings. Every extra post from one channel is penalized. Other people's digests are dropped early, spotted by "week in review" headlines, bullet share, typical sections and a date in the title. **What it does:** - Collected every 30 minutes: Posts pile up all day, the summary is ready by 08:00. - Your own time: Change the delivery time in settings. - Size follows the day: From 10 to 40 posts, depending on how many strong ones came in. - Link to the source: Every item has a "Read →" link to the original post. - Mini App in Telegram: Summary feed, channels, pause a channel in one tap. - Mac app: Native, in SwiftUI: summaries, channels and the collection schedule. **Where to see it:** [suma.evdokimov.ai (RU)](https://suma.evdokimov.ai), [Telegram bot (RU)](https://t.me/mysuma_bot) ### Surkov Brewery - Brewery · 2026 - Case page: https://evdokimov.ai/en/works/pivzavod/ A brewery system that counts beer stock by style and says what to brew and by which day **The problem:** Forty years without brewing. Now they can't keep count. An Arkhangelsk brewery restarted in March. The system collects sales and shipments, counts beer in the tanks and shows six weeks ahead when to brew each style. **What was done:** Built it with AI agents in just over a month: mapping the brewhouse, the calculation, screens for three roles in the brewery's brand book and training videos for the team. **The hard part:** - You brew a month before the tap runs dry. Lager matures in the tank for a month after brewing. So the system counts stock ahead and shows the last day you can still brew with a safety margin. - Only eleven tanks. Brew too much and there's no room for the style you need. Along with the brew day the system picks a free tank and shows which ones free up this week. - The brewer has to see why. No neural networks inside, only rules and plain arithmetic. Any number opens up: beer in tanks and serving tanks, daily outflow and how many days it lasts. **What it does:** - Till and 1C on their own: Sales from the till and shipments from 1C arrive every morning. - Six-week calendar: Which beer runs out where and the last day to brew it. - Manager's screen: What to decide today, free tanks and a «Schedule brew» button. - Brewer's tablet: Shift tasks, transfer marks and a morning tank check. - Owner's summary: One screen: is there enough beer of every style. - In the brewery's brand book: Their colours, fonts and logo, so the system looks like their menu. ### MTS - Telecom · 2026 · live - Case page: https://evdokimov.ai/en/works/mts/ An AI tool for the company's different divisions: edits texts to the editorial policy (Tone of Voice) and explains every edit **The problem:** Fixes the text and explains every edit. A Design Center employee pastes a push notification, app screen or email and gets it rewritten to MTS editorial policy and Tone of Voice. Each edit cites a company rule. **What was done:** Led the project and built the product: a rule base drawn from the company's own documents, the checking logic, a quality test bench, the interface and staff sign-in. **The hard part:** - An explanation can't be made up. Testers caught the model inventing reasons. In the new version the rule base has a second layer: edits follow condensed rules, each explained by a verbatim quote from company documents. - The company's rules, one to one. The rule base mirrors the editorial policy and Tone of Voice section by section. A script, not a person, builds it from the source documents, so it can't drift. - Every version sits an exam. Every new version runs through a bench of 125 texts. It ships only if it does no worse than the previous one. We rolled back even obvious improvements when results dropped. **What it does:** - Any text: A push, an app screen, an error, a support reply, a long read. - Editorial policy and Tone of Voice: The company's own rules, not generic advice about good style. - Each edit explained: What changed and which rule applied. - Quality bench: 125 texts test every new version before it ships. - Corporate email sign-in: Only employees with an @mts.ru address. ### AlfaStrakhovanie OMS - Insurance · 2025 - Case page: https://evdokimov.ai/en/works/alfa/ A Telegram bot for expecting and new mothers with a vaccination calendar, doctor appointments and AI photos of the baby **The problem:** An insurance company moms come to on their own. "Definitely, Mama-Super!" by AlfaStrakhovanie-OMS: a Telegram helper for pregnant women and moms. 24/7 advice, vaccinations, doctor booking under compulsory health insurance (OMS) and AI baby photos. **What was done:** Made with the team, developed together with Continental. **The hard part:** - Nobody wants to read about OMS. The client's brief: engage women with OMS without pushy advertising. The answer: not an insurance ad but a helper that takes on the routine of pregnancy and a baby's first years. - Medicine leaves no room for invention. Advice on pregnancy and babies relies on expert-checked information and Russian research. The assistant remembers each mom's history and answers with her situation in mind. - Moms need more than advice. In-depth interviews showed what matters in this period: keeping memories and just being yourself. So inside there are magic AI photos, video albums and a game with rewards. **What it does:** - 24/7 helper: Remembers each mom's history and answers with her situation in mind. - Vaccination calendar: Syncs with your phone and reminds you of vaccination dates. - OMS doctor booking: One tap, plus help if there's a conflict with the clinic. - Checklists: From "what to pack for the hospital" to the daily routine. - Magic photos: AI turns baby photos into shots in different styles. - Game and momcoins: Points for actions can be swapped for gifts. **Where to see it:** [Telegram bot (RU)](https://t.me/o_mama_super_bot), [Web version mama-super.ru (RU)](https://mama-super.ru), [Channel for moms (RU)](https://t.me/o_mama_super), [Case on Sostav (RU)](https://www.sostav.ru/publication/alfastrakhovanie-oms-tbtbo-i-continental-vypustili-77968.html), [Client press release on RIA (RU)](https://ria.ru/20250602/alfastrakhovanie-2020537979.html), [Digital Leaders 2025 winners (RU)](https://digital-leaders.online/laureatyi/laureates-2025), [Case video (RU)](https://vimeo.com/1116137289) ### Setters - Agency · 2025 · delivered - Case page: https://evdokimov.ai/en/works/setters/ An AI case editor for an advertising agency: a manager dictates the project by voice, and the bot writes the case in the agency's style in 10 minutes **The problem:** ChatGPT writes like everyone. This editor writes like ((SETTERS)). A manager dictates the project facts in a Telegram voice note. Ten minutes later a case comes back with structure, takeaways and the agency's tone. Time per case fell 80%. **What was done:** Designed and built the editor in a month: a breakdown of the agency's cases, a style playbook, a chain of agents, a Telegram bot. **The hard part:** - ChatGPT doesn't write like an agency. General models produce smooth text with no character. We broke down every agency case and draft, then derived the ideal case structure and tone: open on a vivid fact, takeaways, hook phrases. - One model can't do both jobs. One model holds structure better, another writes the brand's tone better. So I built a chain of agents: GPT-4o lays out the facts, Claude Sonnet writes in the agency's voice. - No team to spare for case studies. Commercial projects always come first. So there's one input: a Telegram voice note. Whisper transcribes it, n8n runs the chain, and the brand team gets text in 10 minutes. **What it does:** - Voice instead of a brief: A Telegram voice note, no calls, no transcripts. - The agency playbook: Case structure, takeaways, hook phrases and the ((SETTERS)) tone. - Two models in tandem: One lays out the facts, the other writes in the brand's tone. - Memory: Remembers the agency's cases and its phrasing. - Cases and proposals: Writes commercial proposals as well as case studies. - A human at the finish: An agency editor checks the facts and lays out the final text. **Where to see it:** [Case on Sostav (RU)](https://www.sostav.ru/publication/kejs-alekseya-evdokimova-i-setters-77801.html), [setters.agency (RU)](https://www.setters.agency) ### IMI - Education · 2025 · live - Case page: https://evdokimov.ai/en/works/imi/ An AI bot in Telegram for musicians: finds festivals, labels and specialists, writes press releases, breaks down streaming statistics and puts together a release plan **The problem:** A music industry bot that doesn't make things up. "IMI Base" is trained on materials from the Institute for Music Initiatives: journal, specialists, labels, festivals, jobs. Every answer links to its source, free on Telegram. **What was done:** Spent six months building the bot: the IMI knowledge base, release and stats agents, premium payments, launch. **The hard part:** - An industry where every number is disputed. Models confidently invent moderation times and distribution prices. The bot answers only from IMI's five knowledge bases and catches itself on words like "usually" and "roughly". - The link has to be real. The bot takes sources only from the index of IMI articles. If it can't find confirmation, it gives no link: honesty beats quantity. - Names carry legal risk. Artists, labels, venues: every time one is mentioned, the bot checks Russia's foreign agents registry and adds the mandatory label. **What it does:** - Open calls and events: Open calls and a weekly events calendar. - Labels and media: 100+ labels with contacts, 62 media outlets with pitching rules. - Specialists: From producers to lawyers, with rates and portfolios. - Release plan: Asks about your release and builds a promotion strategy. - Streaming breakdown: A stats screenshot or spreadsheet in, a breakdown of the numbers out. - Live experts: When the base has no answer, it calls in the Chat Krug community. **Where to see it:** ["IMI Base" bot on Telegram (RU)](https://t.me/baza_imi_bot), [Institute for Music Initiatives (RU)](https://i-m-i.ru) ### DeepWedding - Events · 2026 - Case page: https://evdokimov.ai/en/works/deepwedding/ An AI bot that puts together a wedding lookbook for a couple in the brand's collection dresses in one conversation **The problem:** A roughly 45-minute chat instead of five hours. The bot walks the bride through five blocks of questions, hears what really matters, and builds the concept in her words: ten scenes and a lookbook in the collection's dresses. **What was done:** Built the whole bot: the conversation script, dress matching, scene generation, the lookbook and the legal layer. **The hard part:** - In her words, not a questionnaire. After every answer the bot says back what it heard and asks: "Does this resonate?" Until the bride recognizes herself, we don't move on. The concept is built from her phrases. - The real dress, not a lookalike. The 31-look collection is tagged along eight axes. The bot picks three dresses for the concept, a main one, an alternative and a bold one, and describes each in her language. - One bride in every scene. Photos of the chosen dress are the bride's only reference, and couple scenes inherit it down the chain. Other dresses are kept out, otherwise the silhouette gets averaged. **What it does:** - Voice or text: Answer with voice notes and take breaks, everything is saved. - Three concepts to choose from: Pick one, refine it, get a final version of 8,000 characters. - A 21-page lookbook: The meaning of the wedding, ten scenes, the bride's and groom's looks. - Dresses from the collection: Three looks written into her concept, plus booking a fitting. - Gentle follow-up: Check-ins on days 3, 7 and 14, with no new generations. - 152-FZ consent: Two-step consent under Russian data law, text version logged. ### HR monitoring - HR · 2025 - Case page: https://evdokimov.ai/en/works/hr/ For an HR outsourcing firm in Poland: checks nine kinds of employee deadlines daily and warns the HR officer in advance **The problem:** The spreadsheet emails HR by itself when a deadline nears. Every day an n8n workflow at a Polish HR outsourcer reads client employee records, checks nine deadline types and blank dates, and emails HR and the client 30 days ahead. **What was done:** Built the whole workflow: deadline checks, emails to the HR officer and the client, the send history. **The hard part:** - Each contract has its own deadlines. Medical checks and BHP safety training apply only to employment contracts, a mandate only to board members, and open-ended contracts aren't checked at all. The workflow knows what applies to whom. - Dates are written any which way. Dates come as 05.10.2026, 10/05/2026 or "czas nieokreślony". The workflow parses every format, and an empty or impossible date like February 31 gets a separate line. - Remember what was already sent. The send history lives in the spreadsheet itself. A deadline under a week out is sent daily, under two weeks weekly, further out less often. HR doesn't get buried in repeats. **What it does:** - Nine kinds of deadlines: Contract, oświadczenie, legal stay, medical check, BHP and four more. - A 30-day window: Four urgency levels: one, two or three weeks, or a month. - Email to HR: All their employees in one email, from urgent to calm. - Email to the client: Separately for each employee, if the client has its own mailbox. - Overdue and empty dates: Go first in the email, along with the most urgent. - Leavers and duplicates: The workflow skips leavers and resets history for duplicate rows. ### DARS - Real estate · 2025 · delivered - Case page: https://evdokimov.ai/en/works/dars/ A New Year video for a developer, made with ASTRA PRODUCTION: a house spirit decorates an apartment that does not exist yet **The problem:** No apartments yet. New Year's already arrived in them. Developer DARS's New Year video: a yarn house spirit waves a pencil to decorate the apartment, ornaments fly onto the tree, a family walks in. No shoot, 15 seconds, 4K. **What was done:** Made the video with ASTRA PRODUCTION: the character, scenes, animation and the 4K master. **The hard part:** - One house spirit in every shot. A generator draws the character slightly differently each time. The yarn on its head, the orange shirt and the bag with the DARS logo had to match from wide shot to close-up. - Two passes at one scene. The first pass was a warm living room with string lights, the second bright interiors. Over twenty eight-second scenes. The best from the second pass made the cut. - It has to read as real footage. Soft light, bokeh from the string lights, carpet pile and yarn texture. Scenes were generated at 1920 px and upscaled to a 4K master. **What it does:** - Character: A yarn house spirit with a branded DARS bag. - Pencil magic: A pencil instead of a magic wand, sparks and flying ornaments. - A family at the door: At the end, new residents walk into the apartment. - Ready for a series: Two more scripts: new apartment keys and a flight on a lettuce leaf. - Three tracks to choose from: We tried three winter music themes under the video. - 15 seconds: Ad length, 4K master at 24 fps. ### Enotik bot - Architecture · 2026 - Case page: https://evdokimov.ai/en/works/enotik/ A project manager in Telegram for an architecture studio: runs about ten projects, with stages, payments and agreements **The problem:** Seven chats per project, folded into one raccoon. Enotik, a raccoon bot in Telegram, runs an architecture studio's projects: stages, decisions, payments, contracts. The team talks to it like a colleague and each morning learns what's on fire. **What was done:** Built it all: the bot and its 26 tools, the dashboard, the morning digest, PDF intake and contracts. **The hard part:** - Listening to chats didn't work. Version one read the work chats by itself, trying to guess what was happening. It flopped. Version two is a command center: the studio writes, the bot acts, rarely asking back. - Code counts the money, not the model. In the digest the model writes only the status, in words. Amounts, profit and flags like "stuck, target 5 days" come from the database, so bot and dashboard numbers match. - Raw talk into structure. A one-minute voice note about three projects splits into separate records: stage, decision, agreement, payment. The model has 26 tools, from a .docx contract to parsing a product link. **What it does:** - 12 stages with target times: From measurements to design supervision, templates are edited on the dashboard. - Morning digest: At 09:00 Moscow time, changed in settings with no restart. - Voice notes: One message updates several projects at once. - PDFs filed by project: An invoice, quote or completion act finds its project on its own, scans included. - .docx contracts: From the studio's template, the 80% advance and 20% balance calculated automatically. - Payments and settlements: Income, supplier commissions, contractor costs and which partner owes whom. ### Hardware - Prototypes · 2026 - Case page: https://evdokimov.ai/en/objects/ When the object I need is not for sale, I design it myself: five objects so far. ### Lab - Own project · every morning - Case page: https://evdokimov.ai/en/lab/ Every morning I break down one AI technique from fresh videos. ### LTUNEL - Own product · closed - Case page: https://evdokimov.ai/en/works/ltunel/ A members club for the internet as it used to be, 1,300+ members: a Telegram concierge for entry, membership and setup. **The problem:** The internet the way it was ten years ago. LTUNEL was conceived as a private club: entry by invitation, a concierge in Telegram, a Wes Anderson-style scene for every step. 1,300+ people joined; the club is now closed. **What was done:** Came up with it and built it myself: brand book and tone, the concierge bot, membership and payments, infrastructure and support. **The hard part:** - Atmosphere without shields or neon. No locks, globes or glitch. Every bot screen is a Wes Anderson-style photo: a hand holding a curtain, a key on a brass tray, an hourglass on velvet. - A concierge who doesn't sell. The bot talks like the person at the front desk: calm, a little ironic, no exclamation marks or discounts. It explains rather than persuades. The club's vocabulary: join, membership, access. - Grow a little at a time. Entry by a member's invitation or by application. People were admitted in small batches, "so the infrastructure could handle the load", and that's how 1,300+ people joined. **What it does:** - Greeted people at the door: An invitation from a friend, or an application reviewed by hand. - Ran membership: Payment, renewal and a reminder when the term was running out. - First steps: Pick your device and the concierge walks you through, with pictures. - Invite a friend: Invite a friend, and when they join you get a month free. - Gift a membership: A membership can be sent to a friend as a gift. - Help: "It happens, let's sort it out": simple steps first, then a live reply. ### Alfa marketing training - Training · 2025 The marketing team learned to make creative and copy with neural networks in-house. ### Glucose sensor - Health · 2026 - Case page: https://evdokimov.ai/en/works/glucose/ A personal pipeline over sensor, sleep and food data with pattern analysis in a report. **The problem:** Any sensor shows the sugar curve. None shows why. 16 days on myself: plate photos to a bot, a smartwatch, a glucose reading every 5 minutes. The system merges it all into a report explaining what spiked my sugar. **What was done:** Built it for myself over a weekend: a food bot, data collection from the watch and the sensor, analytics, a PDF report and the analysis prompt. **The hard part:** - Food without a diary. I'd quit a manual food diary in three days. So I photograph my plate in Telegram: the model recognizes the food, estimates the weight, counts calories and macros and logs it. - Three streams on one timeline. The sensor sends sugar every 5 minutes, the watch sends heart rate, HRV, steps and sleep, the bot sends food. A script aligns them and measures each meal's rise, peak and recovery. - Analysis, not a retelling. The prompt forbids the model to repeat numbers: "glucose 8.0" tells you nothing. It compares meals with each other, accounts for sensor lag and looks for what contradicts the tables. **What it does:** - A photo instead of a diary: Snap your plate, the bot returns foods, weight, calories and meal time. - Response to every meal: Rise, peak, speed and how many minutes sugar took to return to baseline. - Forecast versus fact: What the table's glycemic load promised and what the sensor showed. - Nights on their own: Night profile, morning rise without food, dips below normal. - Heat map: Average for each hour of each day: high-sugar hours stand out. - Day score: One number from five metrics, with its trend over the whole period. **Where to see it:** [Post about the system on my channel (RU)](https://t.me/ltunevibe/1212) ### Volgograd of the Future - Exhibition · 2024 - Case page: https://evdokimov.ai/en/works/volgograd/ A series of works about the city of the future in a real exhibition. **The problem:** My city many centuries from now. No catastrophe. Three A1 posters for the "Volgograd of the future" show: a positive dystopia where the river station's "puck" and the Stalingrad Panorama are monuments, and the city grew into nature. **What was done:** Came up with the series and made three pieces: concept, generation, finishing for A1 print and the author's text for the hall. **The hard part:** - The buildings had to be recognizable. The posters hung in Volgograd for locals. So the silhouettes of the "puck" and the Panorama rotunda read at first glance. Only time, light and the nature around them change. - Generating at A1 size. An A1 poster at 300 dpi is 7017 by 9933 pixels. No generator outputs that much, so each piece was upscaled and retouched to print quality. - A future without gloss. Instead of flying cars, quiet scenes: two elderly people on the embankment looking at the city, a figure in the grass by the Panorama. The person is small, time is big. **What it does:** - Three A1 posters: The embankment, the river station "puck" and the Stalingrad Battle Panorama. - The third hall: The "Through Designers' Eyes" exhibition at Rechport, third hall, alongside other authors. - A social media series: The carousel "Volgograd. 386 Years Later" with close-ups. **Where to see it:** [Article about the exhibition on V1.ru (RU)](https://v1.ru/text/gorod/2024/11/03/74195516/) ### Arkan - Music · 2024–2025 - Case page: https://evdokimov.ai/en/works/arkan/ An artist built on the full neural stack: image, videos, first live show. **The problem:** He doesn't exist. He performed for a live audience. Arkan is made entirely with neural networks: character, music, 22 tarot cards as the plot of his life, four streaming tracks and videos for his first live show. **What was done:** Came up with him and built him together with Ilya Menzeleev: character, tracks, visuals, videos and a talk at the festival. **The hard part:** - Character before the first note. We built the personality before the music: the childlike sincerity of school essays, Rick Rubin's wisdom and Kanye's nerve. Lyrics, his way of talking and his arc grew from that. - A sound that's recognizably his. Any AI makes a smooth track in a minute. We wanted our own soul sound, so each track is built from hundreds of small generations, a lab exercise in music production. - Music videos in three days. Three days before the fest we realized the talk needed a video for every track. One video usually takes a week. Auto-prompting, over a thousand images, a tenth made the cut. **What it does:** - Four tracks: "Spirit of My City", "Say", "Stand Behind Me", "Farewell" on streaming. - A 22-card deck: The story is bound by the deck: one card, one step on his path. - Music videos: Midjourney, Runway, Kling, Luma and Higgsfield in one pipeline. - A live stage: Headliner of the Maker Fest education festival, with videos inside the talk. - A podcast on the process: How we built a personality and why musicians fear projects like this. - His own sign: A black swan runs through the covers, the cards and the logo. **Where to see it:** [Instagram @arkanportal (RU)](https://www.instagram.com/arkanportal/), [Telegram @arkanportal (RU)](https://t.me/arkanportal), [Yandex Music](https://music.yandex.ru/artist/18783229), [Apple Music](https://music.apple.com/ru/artist/arkan/1746534364), [Spotify · "Spirit of My City"](https://open.spotify.com/album/7gEU1GvpJoywVfkB6NYZjT), [YouTube Music](https://music.youtube.com/playlist?list=OLAK5uy_mhi53sKisLHmCIikPlpcMbRt1GRMmB77A), [Music videos on YouTube](https://www.youtube.com/playlist?list=PLZ5w5hmoANPPGeoct5YdV9P1LGSWVPu3T), [Podcast on YouTube (RU)](https://youtu.be/ytdW8xvEHeU), [Podcast on Rutube (RU)](https://rutube.ru/video/827ee84f7b6124a786c9282ec0c9df5c/) ### Trends bot - Tools · 2026 - Case page: https://evdokimov.ai/en/works/trends/ A Telegram bot and mini app: ask across three hundred industry reports, get an answer with numbers and the report's name. **The problem:** A crystal ball that's read three hundred trend reports. TRENDY is a Telegram bot and mini app built on industry reports from 2025 and 2026: each answer gives takeaways with numbers and names the report. **What was done:** Built it myself: collected and tagged the report corpus, then built the search, prompts, bot and Mini App. **The hard part:** - I had to build the corpus. Since late 2024 I've been collecting, structuring and digitizing reports. I wrote a description and category for each, so every answer shows which report a number came from. - Takeaway first, number as support. Two models take turns: a fast one pulls facts from search results, a strong one interprets them. The prompt bans the obvious, filler and any number missing from the reports. - "More" without repeats. On a follow-up the model gets everything it already said, and all of it is off limits: no repeating trends, numbers or companies. Nothing left to add? Then it says so. **What it does:** - 12 topics, one tap: AI, media, health, finance, retail, food and six more. - "More", up to three times: Each follow-up brings different facts and different companies. - The report under the takeaway: The source's name sits at the bottom of every trend card. - Related topics: Three next questions under each answer, to dig deeper. - Report list: 180 reports from the 2026 edition by category, with descriptions and search. - History and copy: Past questions are saved, an answer copies in one tap. **Where to see it:** [Post on the TRENDY2025 launch (RU)](https://t.me/ltunevibe/845), [Post on the TRENDY26 Mini App (RU)](https://t.me/ltunevibe/1137) ### Unspark - Tools · 2026 - Case page: https://evdokimov.ai/en/works/unspark/ A Mac app that locally cleans the corner of my own generated images, no cloud. **The problem:** The star left the corner. The picture stayed whole. My Gemini and Nano Banana generations come with a ✦ mark in the corner. Unspark finds it and repaints the corner with the local LaMa model, right on the Mac, offline. **What was done:** Built it for myself end to end: the mark detector, the LaMa wrapper, the interface and the Mac app build. **The hard part:** - Find the mark on any background. The star can be gray on dark fabric or near white on a pale floor. The detector hunts the ✦ shape, not color, in the lower-right third after a background-muting filter. - Repaint, don't smear. LaMa sees the whole image, so it continues the terrain, shoreline or background noise under the mark. At 100% zoom there's no seam. - Offline on an ordinary Mac. The model and weights ship inside the app, no network needed. It runs on the Apple Silicon CPU, no graphics card: 1024–1280 px in 7–12 seconds. **What it does:** - File or folder: Drop one image or a whole folder, subfolders included. - Original untouched: The result lands alongside with an _unsparked suffix, the original unchanged. - Safety net: If the detector isn't sure about the mark, it fills a fixed corner. - Only the corner: Everything outside the mark matches the original pixel for pixel. - No network: Model weights inside the app, "Local · Offline" in the status bar. - Path on screen: Processing time, where it saved, and Save and Process more buttons. ### Homeovox - Pharma · 2025 - Case page: https://evdokimov.ai/en/works/homeovox/ For Zdravsiti, TooMuch had bloggers lose their voice mid-video before a Homeovox ad. I made the AI hoarseness from their own recordings. **The problem:** AI can do any voice. Except a hoarse one. TooMuch's campaign for Zdravsiti: mid-video, Sobchak, Petya English and other bloggers lose their voice, a Homeovox ad runs, the voice returns. Two pack sizes sold 183% and 96% more (Sostav). **What was done:** Did the voice part: hoarse clones of nine bloggers' voices, the takes and the mix. The idea and the campaign are by the TooMuch agency. **The hard part:** - AI has no sore throat. ElevenLabs and plugins let you tweak creativity and emotion. There's no setting for hoarseness, and processing on top kills recognizability: you get a robot or just dirty sound. - The rasp came from the bloggers. Each creator recorded two samples: their normal voice and a deliberately hoarse one. The clone was built from the hoarse one, so it rasps in their own timbre, not someone else's. - Two takes sound like two people. Clean and hoarse tracks cut back to back sounded like a second voice over the first. So we picked takes by stability and similarity and built the transition by hand in Audition. **What it does:** - Nine voices: Sobchak, Petya English, "Guitar from Scratch", "Poslezavtra", Geography Trophy and others. - 72 takes: Hoarse clones with different stability, similarity and pace. - 44 Audition sessions: The shift from a clean voice to a hoarse one, built by hand. - The voice comes back: After the Zdravsiti QR card, the video goes on in a normal voice. - Recognizable: Viewers hear their own blogger, just with a cold. - Two months on air: The campaign ran from April 15 to June 15, 2025. **Where to see it:** [Case on Sostav (RU)](https://www.sostav.ru/publication/kejs-toomuch-i-zdravsiti-povyshenie-prodazh-na-183-s-pomoshchyu-nejrosetej-i-kreativnoj-integratsii-77973.html), [Sobchak integration (RU)](https://www.youtube.com/watch?v=reGBetvhrps&t=1286s), [Petya English integration (RU)](https://www.youtube.com/watch?v=-kO3m7kLcB8&t=1034s), ["Guitar from Scratch" integration (RU)](https://www.youtube.com/watch?v=5y9ktXth1p4&t=444s) ### UMALU - Brand · 2026 - Case page: https://evdokimov.ai/en/works/umalu/ A beach umbrella sharing brand in Florida: logo, posters, canopy layout for the factory. **The problem:** A brand whose logo opens up over the beach. UMALU is beach umbrella sharing in Florida: the umbrella already stands in the sand, your phone opens it. The brand is meant to turn every open canopy into a sign. **What was done:** Built the whole brand: platform, mark and lettering, color, typefaces, icons, photo style, posters, social, website and factory files. **The hard part:** - Standing out against blue. A beach is blue sky, blue water and hundreds of other umbrellas. Every exploration landed on the same coral, PANTONE 171 C. You can spot it even from a hotel balcony. - A sun people hide from. The mark is a sun over the letter "a". It has eight rays, like the canopy's eight panels, so an open umbrella reads as the logo on its own, even without lettering. - Fabric forgives nothing. The canopy geometry came from the factory's template. The lettering never crosses a seam, text is outlined, and the file prints without fonts or anyone else's settings. **What it does:** - Brand platform: Positioning, voice and approved phrases in English. - Mark and lettering: The sun over the "a", vector lettering, an app icon. - System: PANTONE colors, typefaces, 37 icons, supergraphics and a photo style. - Posters and outdoor: Four posters, a backlit city poster, an aerial banner over the beach. - The umalu.app site: One page for a phone on the beach: how to open an umbrella. - Factory files: A 100 by 100 cm canopy layout and a print spec. ## Background ### Experience - More than 15 years in IT and marketing - 13+ corporate AI projects and 50+ AI rollouts ### Clients Key clients of the last 15 years. TSEKH, Kearney, Zewa, Ozon, MTS, Burger King, McKinsey & Company, Ostankino, Körting, MR Group, OLIMPBET, EAPTEKA, SIBUR, Sandoz, Soyuzmultfilm, KT Team. ### Teaching - Taught the «AI in Communications and Branding» course in the Integrated Communications master's program, HSE University School of Communications, 2025 - Executive training: 50+ top managers have taken it - Tsekh: an in-house vibe coding workshop for the studio's producers; in an hour they built a Telegram bot - Skillbox, 2023: expert for two longreads on neural networks for the music faculty - Speaker at the IKRA school ### Career - TBTBO Brand Mastering, 2017–2025: co-founder and creative producer of the branding agency - Post Modern Technology, 2012–2025: marketing director for the Medialog medical information system, which runs in more than 1,000 medical institutions - Optima group of companies, 2009–2012: deputy marketing director, a 5,000+ person holding - Microsoft, 2005: Partner Account Manager ### Personal - Author of the «Evdokimov as usual» Telegram channel - L-Tune, music project: albums Stay Tuned (2013) and Sunshine Album (2010) ## In the press - [How to cut case-study production time by 80% with an AI assistant](https://www.sostav.ru/publication/kejs-alekseya-evdokimova-i-setters-77801.html) — Sostav.ru (Russian) - [The AI that actually does everything for you: why Manus beats ChatGPT](https://incrussia.ru/share/manus-ii/) — Inc Russia (Russian) - [How it feels: why we are trusting ChatGPT more and more](https://knife.media/pokinula-chat/) — Knife.media (Russian) - [How to teach FLUX.1 to draw your portraits](https://vc.ru/u/2305137-alexey-evdokimov/1482401-kak-nauchit-flux1-sozdavat-vashi-portrety) — VC.ru (Russian) ## Contact - **Email:** alexey@evdokimov.ai - **Telegram (direct):** https://t.me/ltune - **Telegram channel:** https://t.me/ltunevibe - **Website:** https://evdokimov.ai/en