Om Jee Mishra | Senior Software Developer

TYPESCRIPT/JAVASCRIPT (ES6+)/REACT/NEXT.JS (SSR)/TAILWIND CSS/REDUX/RECOIL/NODE.JS/EXPRESS.JS/GRPC/WEBSOCKET/SSE/REST APIS/JWT / OAUTH2/LLMS/RAG PIPELINES/EMBEDDINGS/PROMPT ENGINEERING/MASTRA AI/GOOGLE VERTEX AI/GEMINI LIVE API/PINECONE/VECTOR SEARCH/EVALS/MEMORY MANAGEMENT/POSTGRESQL/MYSQL/MONGODB/REDIS/SQLITE/DRIZZLE ORM/DOCKER/AWS/CI/CD/GITHUB ACTIONS/GIT/LANGFUSE/OPENTELEMETRY/VERCEL AI SDK/ZOD/RAZORPAY/CASHFREE/ELEVENLABS/—/TYPESCRIPT/JAVASCRIPT (ES6+)/REACT/NEXT.JS (SSR)/TAILWIND CSS/REDUX/RECOIL/NODE.JS/EXPRESS.JS/GRPC/WEBSOCKET/SSE/REST APIS/JWT / OAUTH2/LLMS/RAG PIPELINES/EMBEDDINGS/PROMPT ENGINEERING/MASTRA AI/GOOGLE VERTEX AI/GEMINI LIVE API/PINECONE/VECTOR SEARCH/EVALS/MEMORY MANAGEMENT/POSTGRESQL/MYSQL/MONGODB/REDIS/SQLITE/DRIZZLE ORM/DOCKER/AWS/CI/CD/GITHUB ACTIONS/GIT/LANGFUSE/OPENTELEMETRY/VERCEL AI SDK/ZOD/RAZORPAY/CASHFREE/ELEVENLABS/—/

§ 01

Selected work, told as case studies.

3 of 7 · full archive →

YEAR2025 — present

ROLESenior Software Developer · Architect

TEAMHexaHealth · AI platform

SCALE10K+ docs · 50+ live voice sessions

Mastra AIGemini 2.5 FlashGemini Live APIPineconePostgreSQLWebSocketLangfuseOpenTelemetry

01 / SAKSHAM

A multi-agent RAG platform with real-time voice, built for healthcare workflows.

Mastra-orchestrated agents on Gemini 2.5 Flash, a Pinecone vector store across five namespaces, and a bidirectional voice channel that doesn't trip over itself under load.

Problem

Healthcare assistants need accurate answers under 2 seconds against 10K+ indexed documents, plus a real-time voice channel where overlapping responses are not acceptable. Existing patterns either lose latency to retrieval or lose stability the moment voice gets bidirectional.

Architecture

Multi-agent RAG on Mastra AI with Gemini 2.5 Flash. Retrieval runs over Pinecone (768-dim embeddings across 5+ namespaces). Persistent conversational memory lives in PostgreSQL with observational layers and 7 LLM-callable tools (symptom checker, semantic search, appointment booking). Voice streams over WebSocket using Gemini Live API, gated by a 4-state turn-management machine. Langfuse + OpenTelemetry trace every turn end-to-end.

Key decisions

4-state turn-management machine eliminated overlapping-response bugs in voice
Prompt-injection detection at 0.7 threshold; one-time session tokens with 60s TTL
IP-based rate limiting at 3 connections/IP — production-safe defaults
Per-turn tracing in Langfuse + OpenTelemetry; nothing ships without an eval path

Outcome

Sub-2s retrieval over 10K+ healthcare documents. 50+ concurrent voice sessions sustained; voice latency down 40%. Unauthorized access cut 95% via the layered security controls.

<2s

p99 retrieval

10k+

docs indexed

−40%

voice latency

−95%

unauthorized access

Visit live ↗Case details →

02 / POSTMIND AI

Content intelligence for LinkedIn & Twitter — a 6-agent pipeline with a Writing-DNA voice engine.

A production monorepo that compresses the blank-page-to-publish loop without sounding like every other LLM tool. Voice fingerprint, adaptive feedback, and shipping infra that earns its keep.

YEAR2025

ROLEFounder & Engineer

SCALETurborepo · 3 packages · 13 collections

Mastra AINext.js 14TypeScriptMongoDBTurborepoDockerGitHub Actions

Problem

AI writing tools generate fluent but off-voice content, then bloat their backends until shipping becomes a chore. The harder problem is staying on-voice without paying for it in latency or per-request LLM cost.

Architecture

Mastra AI 6-agent pipeline with sequential orchestration (Persona Analyst → Trend Research → Content Generator) generates 5–7 personalized briefs per request. A Writing-DNA engine extracts 15+ quantitative voice metrics deterministically — zero LLM cost, 100% reproducible. An adaptive feedback loop weights 6 implicit behavioral signals with a 14-day recency decay. The Turborepo monorepo ships 3 packages, 40+ shared TS interfaces, 11 RESTful API modules with Swagger, 3-tier rate limiting, and 13 MongoDB collections. Next.js 14 frontend with 10+ pages including real-time chat and an AI co-writing editor.

Key decisions

Voice fingerprint extracted deterministically — 15+ metrics, zero LLM cost
6-agent pipeline orchestrated sequentially; each agent owns one job
Docker image trimmed 2.66 GB → 338 MB (87%) via 4-stage multi-stage builds
GitHub Actions CI/CD with multi-platform builds, SHA-tagged versioning, health-check rollouts

Outcome

Live content-intelligence platform generating 5–7 on-voice briefs per request. Production CI/CD with health-check-gated rollouts and a Docker footprint 87% smaller than the naive build.

agent pipeline

15+

voice metrics

−87%

docker image

MongoDB collections

Visit live ↗Case details →

03 / LIT LAW

AI contract drafting for Indian law — built to ship, scale, and not lose a row.

Backend microservices over gRPC, custom-embedding RAG on Pinecone, real-time collaboration over SSE, and payments wired through two providers at 100% reliability.

Problem

Legal contracting on the web is mostly PDFs and DocuSign. Lit Law needed real-time concurrent editing, AI-drafted clauses backed by retrieval, hard uptime guarantees, and reliable payments — all at the throughput a production legal platform actually faces.

Architecture

Node.js + Express microservices over gRPC. PostgreSQL with Drizzle ORM as the system of record, Redis for caching. 40+ JWT-authed REST APIs. RAG pipelines use custom embeddings on Pinecone. Real-time collaboration (presence, concurrent editing) is delivered over SSE. The Next.js (SSR) frontend uses React, Redux, Tailwind, and Tiptap. Razorpay and Cashfree handle payments.

Key decisions

Custom embeddings over off-the-shelf — better domain fit on Indian legal corpus
SSE for collaboration instead of WebSocket — simpler ops at the scale we needed
95%+ test coverage on the backend; nothing merges without a green eval path
Two payment providers with idempotency keys — Razorpay + Cashfree at 100% reliability

Outcome

99.9% uptime, 2.3× throughput improvement, 48% latency reduction. Scales to 1,500+ concurrent users with 500+ monthly signings across two payment providers.

99.9%

uptime

2.3×

throughput

−48%

RAG latency

500+

signings / mo

Visit live ↗Case details →

YEAR2023 — 2025

ROLEBackend Lead

VOL.1.5k concurrent · 500+ signings/mo

Node.jsExpressPostgreSQLDrizzle ORMgRPCPineconeNext.jsRazorpaySSE

§ 02

How I think about the stack.

grouped by purpose, not popularity

cluster · 01primary · 3 yrs

Frontend

Where most product work lands. Strong on typed React, Next.js SSR, and the state-management discipline that keeps long-running apps honest.

TypeScriptJavaScript (ES6+)ReactNext.js (SSR)Tailwind CSSReduxRecoil

cluster · 02production · daily driver

AI & Agentic systems

Multi-agent RAG, real-time voice, structured prompting, eval harnesses, and the boring 90% that keeps a production LLM honest under load.

LLMsRAG PipelinesEmbeddingsPrompt EngineeringMastra AIGoogle Vertex AIGemini Live APIPineconeVector SearchEvalsMemory Management

cluster · 03primary · 3 yrs

Backend & APIs

Node + Express microservices over gRPC and REST, with WebSocket/SSE for real-time. Comfortable owning a service end-to-end.

Node.jsExpress.jsgRPCWebSocketSSEREST APIsJWT / OAuth2

cluster · 04system of record

Databases

PostgreSQL with Drizzle as the default; Redis for cache; Mongo where document-shape fits; SQLite for the edge cases.

PostgreSQLMySQLMongoDBRedisSQLiteDrizzle ORM

cluster · 05ship discipline

DevOps

Docker, GitHub Actions, multi-platform builds, SHA-tagged versioning, and health-check-gated rollouts. Prefer boring infra over clever infra.

DockerAWSCI/CDGitHub Actions

cluster · 06wired across the stack

Tools & Integrations

Observability via Langfuse + OpenTelemetry. Schema safety with Zod. Voice with ElevenLabs. Payments via Razorpay and Cashfree, idempotent end-to-end.

GitLangfuseOpenTelemetryVercel AI SDKZodRazorpayCashfreeElevenLabs

cluster · 07applied across all

Practice & craft

The skills no resume has a row for. Migration planning, on-call discipline, post-mortems that change behavior, code reviews, and getting things to ship.

Architecture reviewMigration sequencingCode reviewsEval harness designIncident responseMentorship

§ 03

Impact log, reverse-chronological.

3 yrs · 2 companies

Jul 2025 — nowSenior Software Developer

HexaHealth

Gurugram, India · Mastra AI · Gemini 2.5 Flash · Gemini Live API

Architected the Saksham multi-agent RAG platform on Mastra AI with Gemini 2.5 Flash and Pinecone (768-dim embeddings across 5+ namespaces), achieving sub-2s retrieval over 10K+ indexed healthcare documents
Engineered real-time bidirectional voice streaming over WebSocket using Gemini Live API with a 4-state turn-management machine, eliminating overlapping-response bugs and handling 50+ concurrent sessions
Built persistent conversational memory in PostgreSQL with observational memory layers and 7 LLM-callable agentic tools (symptom checker, semantic search, appointment booking)
Integrated end-to-end observability via Langfuse + OpenTelemetry with per-turn tracing for both text and voice channels
Reduced voice latency by 40% while sustaining 50+ concurrent live sessions
Cut unauthorized access attempts by 95% through layered security controls

Outcomes

Retrieval p99<2s

Docs indexed10k+

Voice latency−40%

Unauthorized access−95%

Live voice sessions50+

Jun 2023 — Jul 2025Software Developer

Wasserstoff RJ Innovations Private Limited

Gurugram, India · Node.js · Express.js · PostgreSQL

Engineered the backend for Lit Law (Live) using Node.js, Express.js, PostgreSQL, Redis, and Drizzle ORM in a microservices architecture over gRPC
Built 40+ REST APIs with JWT auth, delivering 99.9% uptime and 2.3× throughput improvement
Implemented RAG pipelines with custom embeddings and Pinecone vector search for legal contract generation
Built real-time collaboration (presence, concurrent editing) via SSE and a responsive SEO-friendly UI with Next.js (SSR), React, Redux, Tailwind, and Tiptap
48% latency reduction on RAG retrieval, scaling to 1,500+ concurrent users with 95%+ test coverage
Shipped Lit Law and Lit Sign to production with multi-party signing and clause negotiation

Outcomes

Uptime99.9%

Throughput2.3×

RAG latency−48%

Concurrent users1.5k+

Monthly signings500+

§ 04

Open source, contributed in the open.

8 merged · 1 project

01 / C.O.R.E.

Your Personal AI OS

An open-source, self-hostable memory backbone — a "personal AI OS" that gives agents like Claude Code, Codex and Cursor a shared, portable memory graph over MCP. I contribute across its stack: MCP tooling, third-party integrations, the self-hosted local-LLM path, and agent-service test coverage.

Repository ↗Live site ↗All my PRs ↗

REPORedPlanetHQ/core

ROLEOpen-source contributor

STARS★ 1.8k

LANGTypeScript

YEAR2026

TypeScriptMCPMastra AIOllamaNext.jsVitestCLI

PRs · all merged

MCP tools shipped

+4.1k

lines contributed

unit tests added

What I shipped

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
Added 62 fast, deterministic unit tests over previously untested agent helpers (action-plan parsing, tool-result truncation, history prep)

Contributions · 8 PRs

Built @core/postman, a read-only Postman integration with personal-API-key auth. Runs an incremental sync every 15 minutes across workspaces, collections, environments, APIs and monitor runs (deduped via lastSyncTime + seen-run IDs), and exposes 14 read-only MCP tools over the Postman REST API.

MCPIntegrationSync

20 filesmerged · May 2026View PR ↗

§ 05

Writing, from the build.

1 post · read all →

01Prompt Caching, Explained: Why Your AI Chatbot Gets Cheaper and Faster the Longer You Talk to ItAI models have no memory, so you resend the whole conversation on every message and the model re-reads it from scratch. Prompt caching is the fix — reuse the computation for the unchanged prefix and cut API costs 50–90%. Here's what's actually being cached, and the attention math behind why it works.Prompt CachingLLMsKV CacheAI SystemsJul 24, 202613 min

All writing →RSS llms.txt

§ 06

Education, where it started.

3 institutions · 2016 — 2024

2020 — 2024 · degree

Bachelor of Technology

GL Bajaj Group of Institutions, Mathura

Computer Science Engineering

Graduated with a CGPA of 8.01
Engaged in multiple project-based learning initiatives
Strengthened expertise in Full Stack Development and AI integration

Mathura, Uttar Pradesh, India

2018 — 2019 · school

Senior Secondary (12th Standard)

Paradise Public School, Mainpuri

Physics, Chemistry, Mathematics (PCM)

Secured 88.8% in CBSE Board Examination
Active participant in science exhibitions and Olympiads

Mainpuri, Uttar Pradesh, India

2016 — 2017 · school

Secondary (10th Standard)

C.R.B Memorial Sr. Sec. School, Mainpuri

Science

Achieved 83.8% in CBSE Board Examination
Won awards for science projects and academic excellence

Mainpuri, Uttar Pradesh, India

§ 07

Contact, open channel.

~24h response · IST hours

If you’re building something that has to ship —
let’s talk.

omjeem558@gmail.com↗

Selected work, told as case studies.

01 / SAKSHAM

A multi-agent RAG platform with real-time voice, built for healthcare workflows.

03 / LIT LAW

AI contract drafting for Indian law — built to ship, scale, and not lose a row.

How I think about the stack.

Frontend

AI & Agentic systems

Backend & APIs

Databases

DevOps

Tools & Integrations

Practice & craft

Impact log, reverse-chronological.

HexaHealth

Wasserstoff RJ Innovations Private Limited

Open source, contributed in the open.

Writing, from the build.

Education, where it started.

Bachelor of Technology

Senior Secondary (12th Standard)

Secondary (10th Standard)

Contact, open channel.

I architect agenticsystems & the infrathat survives production.

I architect agentic
systems & the infra
that survives production.