Telegram Mini App for fitness coaches
How Fit Bot grew from a Google Sheets MVP into an application with coach and client roles, FastAPI, SQLAlchemy and Telegram signature validation.
Each workflow is tied to a working project, its documentation or a public write-up. We show the input, system actions, failures and the point at which a person remains responsible.
Three detailed write-ups with architecture, code, problems and fixes. Each article opens directly on this site.
How Fit Bot grew from a Google Sheets MVP into an application with coach and client roles, FastAPI, SQLAlchemy and Telegram signature validation.
How a ready-made astronaut GLB became part of the interface through custom poses, trajectory controls and smooth Three.js animation.
How to search PDF, Word, Excel and presentation files, return pages and citations, and prevent a model from inventing sources.
A single language-model prompt is not enough for 1C catalogs or tender work. Retrieval is a staged process, and the source and reason for each match remain visible.
Start with the SKU, model, identifier and normalized name. A reliable exact result should not be replaced by a model guess.
When wording differs, embeddings retrieve semantically close items. They produce candidates, not a final answer.
Code and the model compare power, size, material, package contents and other required fields. Any mismatch remains visible.
The web is used only when needed. The result keeps its source URL, while price and product fit remain subject to human review.
The status distinguishes a public write-up from a case that is still being reviewed. A draft is not presented as a published result.
The system reads DOCX, XLSX, PDF and image files, extracts line items and specifications, searches internal sources for matching products and prepares an Excel/PDF draft. Web search is enabled separately, and an external price still requires manager review.
Tested against a regression set of 18 real tender files. We do not claim universal accuracy or measured time savings.
The assistant works in Telegram and on the website, reads prices and available slots from YCLIENTS, collects booking details and hands rescheduling, disputed questions and unfinished requests to a person.
Thirty-eight early checks passed, but a review of 34 live conversations exposed failures that the demo did not show. A CRM event counts as a booking only after separate reconciliation.
Free-form names from Excel and PDF files are matched against a catalog of more than 30,000 items. Semantic search proposes candidates, code verifies fields, and a manager confirms the final match.
The case explains why fuzzy keyword search alone was insufficient and where human review remains necessary.
The agent answers from the product record and conversation history, verifies data in MoySklad and forwards order events to a Telegram hub. A manager can continue the same conversation with the full context.
Automated replies, payment events and human follow-up are separate actions. The bot cannot silently modify an order.
Product photos are cleaned, assembled into listing covers and reviewed before publication. One verified batch contained 1,594 photos for 280 products.
Public processing cost: about $0.10–0.15 per photo. Output quality depends on the source image and is reviewed by a person.
A dispatcher assigns jobs, field workers update statuses through Telegram, and photos are stored separately from operating spreadsheets. The interface highlights incomplete and problematic visits.
Time savings were not measured. The public write-up focuses on the workflow and its platform trade-offs.
Commands for a GSM relay did not require a marketing SMS provider or an alphanumeric sender ID. An older Android phone became a local gateway with a clear delivery log.
This case shows when an AI model is unnecessary and a conventional automation is the better engineering choice.
Choose one repeated workflow and bring real input files or conversations. We identify the data source, acceptable error, agent actions and the point that requires human confirmation.
We use technical cookies (Cloudflare) so the site works correctly. Privacy Policy