Chapter 1 · 2023
The idea
The idea arrived before the technology was ready for it. Clients had accumulated hours of session recordings — valuable material that was hard to return to quickly. To recall a single breakdown you had to rewatch the whole lesson. The gap between "I have a question" and "I have the exact fragment with a quote" took minutes when it should take seconds. It was already clear what the product should be. Only one thing was missing — tools that could make it real without a corporate team and budget.
Chapter 2 · 2023–2025
The wait
While technology caught up with the vision, the work that can't be rushed went on: gathering and preparing the knowledge base. Transcripts of sessions, lectures, practices. Selecting methodologies — not everything, only what had passed through personal experience and proven its result. It was a deliberate choice — to accumulate ready content while waiting for the window, rather than ship something raw. When the window opened, the corpus was already there.
Chapter 3 · January–March 2026
The window opened
By early 2026 everything that had been missing came together: managed file search with native metadata filtering, million-token context windows, data isolation at the retrieval layer. That removed the main technical risk — the privacy of personal sessions was now guaranteed by architecture, not by hoping for the "right prompt". The project's first commit was 24 March 2026. And it wasn't lines of code, but documents: PROJECT.md, REQUIREMENTS.md, ROADMAP.md. Three weeks of design before `import aiogram`.
Chapter 4 · March–May 2026
The build
55 days. 729 commits. 204 releases. The most productive week — 144 commits; the most productive day — 64. There was a pivotal incident along the way: a preview Gemini model started returning quota errors, and within two hours — diagnosis, model comparison on the reference set, rollback to a stable version. Quality dropped by less than 3%, and that hour of firefighting turned into a model-selection methodology the project still uses. By mid-May, Tsalkin AI had stopped being a prototype — it became a daily working tool.
Chapter 5 · May–June 2026
Growth
By mid-May the product worked — then a different stage began: turning a working tool into a product for others. Over the next month — another 761 commits (1,490 total) and 64 releases (268 total). The records moved: the most productive day — 71 commits, the most productive week — 306. But the point wasn't the pace. The first users beyond the inner circle arrived, and the architecture followed them: per-client data isolation at retrieval, transcript segmentation into meaning-level fragments, onboarding for newcomers, a personal goal tracker, an insights store as an entry point, and a metrics dashboard. A feedback loop went live: a 👎 puts the answer into a review queue, and from there the question is added by hand to the set quality is checked against. Tsalkin AI stopped being a personal tool — it became a product that grows together with the people who use it.
Chapter 6 · June–July 2026
The ecosystem
By July the product had matured — and a shift in scale happened: from a single bot to an ecosystem. Another 676 commits (2,166 total), 299 releases, 2,028 tests. But now the thing to count isn't the bot — it's the whole: the brain grew organs. Concierge took its place beside it — a personal client portal, program bots and meeting booking in one system; it draws profile and access from the brain — the shared client profile of the ecosystem. And the brain is learning more: in time, to remember a client as one person — goals, commitments, the path already walked — and to lean on that in its answers; on this foundation a personal knowledge graph and a picture of growth are being laid. Answers found a distinct mentor's voice and began arriving as a stream, no long wait; bookmarks and a course table of contents put the valuable within reach; the same prose now reads equally well in Telegram and on the web. And in the ecosystem's shadow two more organs are already growing. Tsalkin AI stopped being a product — it became an ecosystem: ~180,000 lines of code, still one person.
Chapter 7 · August 2026
The measurement
August added nothing — it measured, for the first time. The brain moved from 0.99.24 to 0.102.0, the corpus grew to 22,747 fragments across 257 sources and 533 hours of transcripts, and the pilot entrance opened: fifteen people, a month of full access, no results yet. But the event of the month was the first measurement of the return loop in the product's history. After an answer the bot asks a question back; of the 344 replies written after one, the system counted 16 as a continuation — 4.7%. Replies that looked like an answer by form: 101, nearly a third. The base is small, 630 questions and 37 people, but a sixfold gap says not that the question is bad — that the product fails to recognise an answer when it gets one. The order of work flipped: recognise the answer first, improve the question second; before the measurement it was the reverse. The same month the showcase was rebuilt: nine claims the product doesn't deliver came off it. No finding is pleasant, and together they beat one more feature shipped — measuring yourself costs more than building on blind.
Your conversations never go into model training. A person improves the product from your ratings: a poor answer goes to review, a good question becomes a yardstick.