Digital résumé
Om Jee Mishra
Senior Software Developer
Multi-agent RAG, real-time voice, and the infrastructure under it — with a heavy bias toward building in the open.
I'm a Senior Software Developer with 3+ years shipping production systems at the intersection of full-stack engineering and applied AI. I architect agentic AI platforms — a text-to-SQL analytics agent whose context engineering cut benchmark tokens 76% and wrong answers to zero, and multi-agent RAG platforms where a single agent runtime drives both streaming chat and realtime voice over the same tools, memory, and knowledge base — grounded in vector search, persistent conversational memory, and live CRM/EHR context, with prompt-injection defense and per-turn LLM evals on top. I care about latency budgets, eval harnesses, and the boring reliability work that lets a team ship without rewriting on Monday — and I do a lot of that work in the open, contributing to getnao's nao and RedPlanetHQ's C.O.R.E.
Open source — built in the open
Open source is where I do my most deliberate engineering — small, well-tested, reviewable PRs against real production codebases. The projects I care about most, with every merged PR one click away:
getnao/nao
Open-source analytics agent — build context with the nao CLI, then chat with your data
An open-source analytics agent teams query from the web, Slack, Teams, Telegram and more. My work fixed how usage and spend are attributed in shared chat threads, added license tracking for org admins, and made live stories discoverable.
- Fixed per-user cost attribution in shared Slack, Teams, Telegram and Mattermost threads: every message was credited to the thread starter, so budget limits blocked the wrong person. Stored the sender on each message with migrations for both SQLite and Postgres and no backfill needed (#1755)
- Built license tracking for admins — organization members, people with access and active users in the last 90 days on the Members settings page, verified end to end on SQLite and Postgres 16 (#1802)
- Made live stories discoverable with a non-modal offer right after the agent creates a story (#1532)
- 6 pull requests, 3 merged — plus bug reports and feature proposals filed upstream
RedPlanetHQ/C.O.R.E.
Your Personal AI OS
A shared memory graph and MCP layer for AI agents. I contributed read-only tools, self-hosting fixes, and a deterministic test suite over previously untested agent internals.
- Shipped 16 read-only MCP tools across two features — a full Postman integration (14 tools) and skills-as-tools (list_skills / get_skill) for coding agents
- Unblocked self-hosted deployments end-to-end: fixed four defects on the local-Ollama path with zero change to the hosted OpenAI / Anthropic / Google providers
- Fixed production bugs across API-key deletion, multi-step chat rendering and Ghost HTML publishing, and closed an audit gap by populating IntegrationCallLog.source for API-triggered actions
- Added 62 fast, deterministic unit tests over previously untested agent helpers (action-plan parsing, tool-result truncation, history prep)
- 9 pull requests, all merged — roughly 4.1k lines contributed
Experience
Senior Software Developer · HexaHealth ↗
Building HexaHealth's applied-AI platform — an internal text-to-SQL analytics agent and HealthGpt, the AI gateway: one agent runtime powering both a streaming text chatbot and a realtime voice assistant over shared tools, memory, RAG, and live patient context.
- Architected an internal text-to-SQL analytics agent (Mastra, Gemini, MCP) over a production MySQL replica — schema sync into a role-scoped knowledge base, tool calling, and context engineering that holds the prompt at ~5k tokens as the schema grows
- Cut its benchmark tokens 76% (874k → 205k), cost 63% and wrong answers 2 → 0 with verified SQL measures and stored-procedure calls; enforced RBAC with a SQL validator and a KILL QUERY backstop, backed by 650 tests
- Architected HealthGpt, the AI gateway — one Mastra agent runtime powering both a streaming text chatbot (SSE) and a realtime voice assistant (WebSocket ↔ Gemini Live, VAD barge-in) over shared tools, memory, and RAG
- Built namespaced Pinecone retrieval over treatments, hospitals, doctors, and uploaded medical PDFs, hydrated each conversation with live CRM/EHR context via LLM summarization, and validated booking server-side so it never fabricates a lead
- Cut a 2–3s voice greeting stall (asynchronous function calling), 1–1.7s off short turns (gated prompt-injection detector) and a ~3s RAG cold start; fixed OTLP exports that broke up to a quarter of Langfuse traces; sampled LLM-as-judge medical scorers on both channels
Software Developer · Wasserstoff RJ Innovations ↗
Owned the backend for Litt Law — microservices, RAG pipelines, real-time collaboration, payments and an AI-chat credit system in legal-tech.
- Built 40+ REST APIs over a gRPC microservices architecture (Node/Express, PostgreSQL, Redis, Drizzle) at 99.9% uptime and 2.3× throughput, monitored with Prometheus and Grafana
- Shipped RAG for legal contract generation with custom embeddings + Pinecone — 48% retrieval latency reduction, scaling to 1,500+ concurrent users
- Delivered real-time collaboration (presence, concurrent editing) over SSE with a Next.js SSR front end
- Integrated Razorpay/Cashfree with 100% reliability across 500+ monthly signings and built a usage-based credit system for AI chat; migrated 8,000+ posts across CMSs with zero data loss
Stack
Writing
I write about the systems I ship — with a heavy focus on open source and the infrastructure decisions that decide whether it survives production.
Education
GL Bajaj Group of Institutions
B.Tech, Computer Science Engineering