I'm 16, built a Nigerian Tax AI agent in one week, here's everything that went wrong

I just shipped Naija Tax AI — a RAG-powered Telegram bot
that answers questions about Nigeria's 2025 tax reform
with citations from the actual law documents.

The stack: n8n (self-hosted on Render) + Supabase pgvector +
Google Gemini embeddings + LangChain + Telegram Bot API.

It works. But getting there involved 12 production bugs including
an encryption key getting rotated mid-build and destroying all my
credentials, Gemini silently returning empty embeddings for weeks
before I realised the model wasn't supported on my API key, and
a shared memory bug where all users were reading each other's
conversation history.

The lessons that actually mattered:
— Chunk size for legal docs needs to be 3x what you'd use for
general text because tax sections cross-reference each other constantly
— Top K matters more than chunk tuning for retrieval quality
— Tool descriptions on AI Agent nodes are more important than
system prompts — wrong description means the agent never calls
the tool at all

1,037 chunks ingested. Bot is live. Zero users so far (working on that).

GitHub: github.com/philipakintola01-sys/naija-tax-ai
Bot: t.me/NIJATAXAIBOT
also looking for scaling and team opportunities
Happy to answer anything about the build.