Telegram Bot, Web, B2B API · In-house product
Personalytic — a personality profile from real speech
Task
Personality tests hand out a label based on self-report. The aim is to see stable patterns in how a person actually writes — with privacy built into the architecture.
What we built
- A Telegram bot analyses group chats only with participants’ consent: about 50 features per message → about 30 traits with confidence and coverage.
- An adaptive 218-question questionnaire, reports from Claude, comparison and compatibility, profile import by code or QR, Telegram Stars.
- A web dashboard and a Mini App.
- A B2B API for profiling support conversations: 73% accuracy overall, 100% on abuse, 92% on escalation.
- A brand platform: from a “personality type” to an observable pattern, descriptive wording instead of evaluative.
Engineering decisions
- Our own morphological engine (79 dictionaries): phrases with skipped words, clause boundaries, negation; changes verified on 5,000 real messages.
- Privacy enforced in the database: schema constraints, cascading deletion on /forget.
- The language model receives only the data a task needs, and every request goes into an audit log.
Stack
TypeScript, Node.js, PostgreSQL, Redis/BullMQ, grammY, Claude API