Production systems for clients and Initializ, the AI agents on top of them, open-source contributions, and tools I built for myself and kept running.
Nestara helps home buyers get loans from many lenders. I’m the top committer on the backend that runs it and I created the CRM its sales team works in every day.
A pydantic-ai agent that must call a typed calculator command before it can answer. Exact-Decimal EMI, eligibility and balance-transfer engines with reverse solvers; rate limits, token budgets and a circuit breaker around the model.
Added WhatsApp as Forge’s fourth channel, in pure Go: QR pairing state machine, closed-by-default admission, a dedup ring as the reply-loop guard, and 16 end-to-end pipeline tests. Also fixed tool entrypoint resolution in the runtime.
A FastMCP server exposing lender-offer, comparison, EMI and email-drafting tools, and a hand-written OpenAI tool-calling loop over MCP Streamable HTTP with one session per lead.
An A2A-protocol agent in Go that finds why warehouse robots stopped: concurrent multi-pass log parsing, token-budgeted LLM batches, findings in Qdrant and MongoDB. RCA turnaround down 60%.
WhatsApp channel adapter (PR #452, +6.2k lines) and a runtime fix for tool entrypoints.
Go · agent runtime ↗initializ/secrets-operatorKubernetes operator that syncs Initializ secrets into clusters. Built from scratch.
Go · Kubernetes ↗loannetwork/PDFFusionMicroservice that merges PDFs and images from S3 presigned URLs, validated with qpdf.
Python · FastAPI ↗ratnesh-maurya/cursor-claude-personas38 role-based persona packs for Claude Code and Cursor.
AI · prompts ↗Developer utilities that run entirely in the browser — nothing you paste leaves the tab.
Habits, expenses, sleep and study in one PWA with its own analytics.
Patient records and treatment tracking for de-addiction centres.
Side-by-side npm package comparison with dependency analysis.