Case study
Sihem: Proactive AI Personal-Mentor & Routine Assistant
Proactive personal-mentor assistant on Telegram (@sihem_ai_bot) and installable PWA (sihem-pwa.pages.dev), accepted into the tech2b incubator in Linz.
Problem
Most habit trackers and todo apps are passive — they wait for manual input and break down as soon as a user's schedule shifts, while raw LLM chats lack persistent structured state, proactive scheduling, and cost-controlled multi-provider reliability.
Role
Solo Developer & Project Lead: full cloud architecture, deterministic nudging engine, pgvector RAG memory, multi-provider LLM routing, Telegram bot, and installable PWA.
Architecture & Pipeline Summary
Deno / TypeScriptSupabase Edge FunctionsPostgreSQL + RLSpgvector RAGLLM Routing (GLM/Gemini/Groq)pg_cronCloudflare MeloTTSTelegram Bot + PWA
1. Activity & Schedule Ingestion
Workouts, medication, study sessions, custom habits, Google Calendar sync, and multimodal meal-photo vision without schema changes.
2. Generators → Arbiter → Composer
Deterministic decision pipeline paired with pgvector long-term RAG memory to trigger adaptive, context-aware, non-spammy check-ins.
3. Multi-Provider LLM & Voice Delivery
GLM, Gemini, and Groq routing with automated failover, Cloudflare MeloTTS voice notes, and delivery via Telegram, PWA, or Alexa Echo.
Approach
- Engineered a serverless architecture on Supabase Edge Functions (Deno/TypeScript) and PostgreSQL with strict Row-Level Security, scheduled via pg_cron.
- Implemented a deterministic generators→arbiter→composer decision engine that evaluates calendar shifts and user habits before synthesizing check-ins.
- Combined pgvector semantic embeddings for long-term user recall with multi-provider LLM routing (GLM/Gemini/Groq) and Cloudflare MeloTTS voice generation.
Results
- Live proactive beta running on Telegram (@sihem_ai_bot) and installable PWA (sihem-pwa.pages.dev) with automated morning briefings and evening reviews.
- Multi-provider routing with automatic failover across GLM, Gemini, and Groq, coupled with granular token cost tracking per user interaction.
- Accepted into the tech2b Upper Austrian incubator (May 2026 – Present) for product validation and user cohort testing in Linz.
Lessons
- Deterministic generators and arbiters prevent LLM hallucination and notification fatigue far better than open-ended prompt loops.
- A unified, generic activity schema eliminates database migrations when introducing new habits, routines, or tracking types.
- Multi-provider failover is mandatory for personal assistants: outages from single LLM APIs immediately erode habit consistency.