Notes from the build.
The hard parts of shipping GenAI — guardrails, agents, cost, evals, and the observability that keeps it honest.
AI drafts, a human decides: running the whole Avni lifecycle on AI
The playbook — scope → design → build → QA → support → orchestrate, a human decision at every stage. (Authored & published on the Avni blog.)
AI drafts, a human decides: running the whole Avni lifecycle on AI
The playbook for taking an NGO from a programme spec to a running, field-ready app — scope, design, build, QA, support, orchestrate — with AI on the typing and a human on every decision.
Cost-capped agents: shipping an AI SDK with a $5 ceiling
A production agent that can't run away with your money — per-session caps, per-turn caps, and a thrash-detector that kills wasted spend.
MCP tool-gates: rules an agent literally cannot skip
Prompts are suggestions. If a rule actually matters, enforce it as a tool the agent must call — and revert anything it does off-script.
Is the model still good? Treating evals as a test, not a hope
Models drift, prompts rot, providers change. A real-LLM eval harness turns "seems fine" into a number you can gate a release on.
AI that reads prod but never writes it
How −60% support resolution time came from an agent with exactly zero write access to production.