For technology partners and grant-makers
An AI system for developing entrepreneurs, built on a practitioner's real experience. A product that by industry standards takes a team and years was built by one person — and keeps growing into an ecosystem.
Quality tools for developing a business today are either expensive or not widely available. The goal is to make them accessible and help owners through the "first mile" — the stage where a notable share of young businesses close. That's both social scale and an open market.
All figures taken on 17 August 2026 — by direct query against production and the repository, not from documents.
One 'brain' — and organs around it. Each takes its own circle of care: some already work, some are on the way.
Client care for programs: a personal portal, bots and meeting booking in one system.
Client care for programs: a personal portal, bots and meeting booking in one system.
| Metric | 18 May | 16 Jun | 11 Jul | 17 Aug | Δ vs previous |
|---|---|---|---|---|---|
| Knowledge fragments in production | ~17 700 | ~21 500 | ~22 400 | 22 747 | +347 |
| Transcript sources (VTT registry) | 125 | 187 | 199 | 257 | +58 |
| Production stores | 3 | 6 | 6 | 6 | — |
| Days of active development | 55 | 84 | 109 | 146 | +37 |
| Commits | 729 | 1 490 | 2 166 | 2 507 | +341 |
| Releases in CHANGELOG | 204 | 268 | 299 | 356 | +57 |
| Automated tests (defs in tests/) | 337 | 1 167 | 2 028 | 2 831 | +803 |
| Python lines — all code, incl. tests | 52 912 | 77 383 | 105 955 | 125 952 | +19 997 |
| Developers | 1 | 1 | 1 | 1 | — |
18 May 16 Jun 11 Jul 17 Aug
Commits. Concierge and new projects join from June — the ecosystem grows organs.
| Component | 18 May | 16 Jun | 11 Jul | 17 Aug | Δ vs previous |
|---|---|---|---|---|---|
| Brain · Mentor | 729 | 1 490 | 2 166 | 2 507 | +341 |
| Concierge | 224 | 374 | 682 | +308 | |
| 2 more projects in development | 165 | 354 | 364 | +10 | |
| One more element of the ecosystem. Soon. | — | ||||
| Ecosystem · total | 729 | 1 879 | 2 894 | 3 553 | +659 |
In lines of code that's ~205,000 lines of Python · 1 developer. A scale that by industry standards takes a team and years — built by one person.
By industry standards: rebuilding a product like this in one year would take 12–20 person-years of work — i.e. a team of 12–20 engineers (backend, ML/AI, DevOps, QA, product). Built by one person in 4.8 months — and still growing.
By industry standards (a simplified COCOMO II–based model adapted for AI/ML): the current engineering codebase (~126 KLOC of core, ~205 KLOC with the ecosystem) corresponds to 12–20 person-years of work. Rebuilding the project in one year would take a team of 12–20: backend, ML/AI engineer, DevOps, QA, product. Built by one person in 4.8 months — efficiency versus the industry baseline on the order of 60–80×: AI-augmented coding, no enterprise overhead, and an extremely focused domain.
Where we are now. On the left — what already works; on the right — what lies ahead. The ecosystem grows through two organs at once: the mentor-brain and the concierge.
Concierge — client care for your programs: a personal web portal, Telegram bots, and meeting booking in one system.
Already shipped
Answers now speak in Maxim's own voice — his manner, his directness, substance without filler.
A detailed answer unfolds on screen as it takes shape — no long silent wait.
Even at peak load the mentor stays with you and delivers the answer, instead of dropping midway.
A valuable answer saves to your personal library in one tap — nothing important gets lost in the chat anymore.
Your program's table of contents is always at hand — a map of the course: where you are and what you've already covered.
Answers arrive cleanly formatted — emphasis, lists and quotes come through intact, not as one flat block of text.
A personal client portal at app.tsalkin.ai — sign in with a link from your email, no password and no messenger needed.
Several programs — Yadro, Praktikum CEO, Praktikum VIP — run on one engine, each with its own branded look.
Book a mentor session in a couple of clicks: pick a free slot in the calendar, and once confirmed you get a Zoom link and a calendar invite in your inbox.
Your whole picture on one page: what you've submitted, what's paid through, your schedule in your own time zone, and every lesson recording in one place.
The 'Owner's Code' diagnostic: an owner answers a set of questions and gets a maturity map of their company — with the option to unpack it in a personal session.
Install the portal on your phone like an app and get push reminders — even on iPhone, even when the messenger is down; email backs up anything critical.
The portal knows its client: it pulls profile and access from the tsalkin.ai 'brain' — the ecosystem's shared client profile.
Ahead
A mentor who doesn't only answer, but reaches out at the right moment — a nudge, a question, a return to what you agreed.
One memory of you: whatever you use across the ecosystem, the mentor knows you as one person, not from scratch each time.
An ever-clearer picture of your growth — not scores, but honest milestones on the way to your goals.
The mentor beyond Telegram — the same voice and the same memory, now on the web.
Mentor memory, the knowledge graph and the growth showcase are in active development.
What's next for Tsalkin AI — planned functionality from the product's working backlog.
From free questions to the programs without manual moderation: hit the limit — get a personalized offer and leave a request right away.
A free self-diagnostic tool for owners — a quick read on the business and its priorities before the first conversation with the bot.
Smart reminders on engagement signals: when a question went unanswered or hit a limit, the bot proposes the next step itself.
Length, style and preset as separate dials. Three tones (standard, tougher, gentler) already ship; profiles will fold them into one setting.
Analytics across the whole path: link click → first question → return → upgrade. To see what actually works instead of guessing.
An eval gate before every release, citation-accuracy control and multi-turn dialogue — answers don't degrade as the audience grows.
Plans, not promises: priorities shift as we work with real users.
Under the hood
Python 3.11, uv. FastAPI for the admin API and internal endpoints, aiogram 3.22 for Telegram. SQLAlchemy 2.0 + Alembic (async ORM, 60 migrations). SQLite: bot.db (users, requests, sessions, enrollments) and content.db (video links). Gemini File Search v2 as the main knowledge backend. APScheduler for cron jobs. structlog in JSON. Docker Compose + deploy.sh. The admin UI is Jinja2 + Alpine.js + Tippy.js — three script tags instead of a build pipeline.
KnowledgeBaseAdapter is a shared protocol with query() and health_check(). Implementations: NotebookLMAdapter and GeminiFileSearchAdapter. Swapping the backend is one env variable. When the external API became unstable, migration took 2 days instead of 2 weeks — because the Q&A logic didn't know which backend was underneath it.
Personal sessions are stored under PART_XXXX codes. Before every LLM stage the codes are swapped for real names in memory only; after the answer, swapped back. Storage stays neutral, the LLM sees human names, nothing extra leaks out. This closes the risk of a model "inventing" a name inside someone else's session.
A privacy-critical pattern: the client_id filter is applied at retrieval (a native Gemini File Search feature), not after. The LLM physically never receives someone else's chunks — isolation doesn't depend on a well-behaved prompt. Meta-queries like "which sessions do I have?" are answered not by the LLM but by a structured query to the registry — no fact hallucination.
A generic CachedRepo with a TTL+LRU cache and async DB fetch. Concrete repos for access grants, tier-parameter overrides, the audit log, corpus documents. The signature matches a future RedisCache — switching is one line. The old YAML readers don't break: migration goes reader by reader with a 30-day fallback. Evolution instead of a big rewrite.
Tier (the billing and quota plane) and program enrollment (the content-access plane) are separated. "MENTEE" used to mean both. Now a user can be CLIENT by tier with enrollments in several programs. This gives linear scalability to 50+ programs with no code changes.
One SDK: embeddings and retrieval inside a single client. Native custom_metadata + filtering at retrieval — the foundation of personal-session privacy isolation. For a solo developer, a separate vector service means an extra bill, an extra SDK, an extra source of night-time alerts. Vendor lock-in is contained by the KnowledgeBaseAdapter abstraction: moving means a new interface implementation, not rewriting the bot.
Development dynamics: 18 May → 17 August 2026.
A watchdog with a marker pattern instead of PagerDuty: a deploy sets a marker — the alert stays silent; an OOM ignores the marker — the alert always fires. Zero false positives across six deploys in a day. CHANGELOG in sync with every commit — it compensates for the lack of code review. All source-of-truth mutations go through an explicit admin click: surface + friction + record beats doing it silently. The paid eval harness runs only on explicit approval.
Opening the engine core is a direction for the project (open-core): what isn't proprietary value can be opened to the community. It's an engineering contribution and transparency for partners. The exact open/closed boundary is being worked out.
An AI product of this complexity — RAG, multi-tier, admin panel, eval harness — would, by industry standards, take a team of 12–20 engineers to rebuild in a year: backend, ML/AI, DevOps, QA, product. Tsalkin AI was built by one person in 146 days.
That's my superpower and my mission: to give today what will become widely available tomorrow. Effective solutions with a fraction of yesterday's resources — using knowledge, experience and technology.
2,507 commits. 356 releases. 2,831 automated tests. 125,952 lines of Python across the repository — 82,280 of them the bot's production code, the rest tests and data prep. One person, under five months. This isn't theory — it's already done. Figures taken on 17 August 2026.
— Maxim Tsalkin
The project is open to grant support and technology partnerships: joint pilots, integrations, infrastructure access. At its core is a rare intersection of business practice, mentoring and in-house technology.
If you represent a fund, an accelerator, an SMB bank or a technology company — reach out directly.