Full Stack Engineer building agentic AI systems
TypeScript · Node.js · React / Next.js · Python · AWS · LLM agents, RAG, MCP and evals
I build production web platforms and the LLM systems that run inside them. Over 5 years I have shipped full-stack products for clients at UST, Teckollab and Emvigo Technologies. My focus now is agents that call real tools safely, retrieval that is measured instead of guessed, and the plumbing (gateways, guardrails, tracing, evals) that makes AI features dependable in production.
Open to Full Stack and AI Engineer roles, remote or on-site in India.
- 6x faster API at 2.7x the traffic. On a multi-portal facility-operations platform (Node.js, MongoDB), traced an APM hotspot where one login triggered ~4,400 database writes, cut it to ~6 and added targeted indexes. Average response time went from 130 ms to 20 ms and CPU from 18% to 2%; slow screens dropped from 1.4 s to 0.06 s.
- 2+ weeks of work in 3 days. Built the platform's review automation engine with a spec-driven, agent-assisted development workflow, reviewing the core paths by hand and verifying the rest with tests.
- Embeddings pipeline on PostgreSQL + pgvector (generation, storage and retrieval) powering personalization and carbon-impact scoring for a sustainability platform for UK SMEs.
- Security hardening: replaced
eval()in a cron scheduler with an allow-list, locked down an open error-logging endpoint and removed hard-coded admin credential seeding. - End-to-end delivery of an athlete-sponsorship marketplace and a video micro-learning platform (React / Next.js, Node.js, Azure DevOps CI/CD, webhooks, analytics dashboards).
AI engineering
| Project | What it is | Stack |
|---|---|---|
| llm-gateway | Self-hostable LLM gateway: model aliases with fallback, retries with jittered backoff, circuit breakers, per-key rate limits and USD budgets, response cache, PII redaction, Prometheus metrics and OpenTelemetry traces, load-tested. | TypeScript, Fastify, Redis, Docker |
| rag-eval-lab | RAG over the Model Context Protocol docs where the eval harness is the product: 2 chunkers x 4 retrievers (BM25, dense, hybrid RRF) x reranker ablation, recall@k / MRR / nDCG, citation and refusal accuracy, and a CI regression gate. | Python, FastAPI, sentence-transformers |
| mini-coding-agent | A coding agent written from scratch (no frameworks) plus a SWE-bench-style harness: 20 tasks with hidden tests, sandboxed file tools, context summarization, failure taxonomy and a trajectory viewer. | Python, pytest |
| telemetry-ops-copilot | Operations agent for an IoT fleet over MCP: six tools, human approval for every write action, step and cost budgets, JSONL traces with a viewer, and a 25-case eval suite that checks zero unapproved writes. | TypeScript, MCP SDK, SQLite |
Platforms and tools
| Project | What it is | Stack |
|---|---|---|
| edge-telemetry-platform | Edge-to-cloud industrial telemetry on AWS: Greengrass v2 at the edge, IoT Core rules, IoT SiteWise asset models and a typed query API, all as infrastructure code. | TypeScript, AWS CDK, Lambda |
| talkies | Narrated product demo videos as code: Playwright scenes, automatic dead-air cutting, text-to-speech voiceover, one synced MP4. | Node.js, Playwright, FFmpeg |
| agentic-ai-wiki | A public, LLM-maintained knowledge base of practitioner best practices for building with agents, researched and updated daily by a scheduled agent in GitHub Actions. | GitHub Actions, Markdown |
| penumbra | Open-source Android drawing app: pressure and tilt brushes, layers with blend modes, generated coloring pages, stroke replay, OpenRaster files. Framework-only, no third-party libraries. | Kotlin, Android |
| Gmail-Auto-Cleanup | Apps Script that keeps a Gmail inbox clean by rule while protecting important mail. My most-forked repo. | JavaScript, Apps Script |
- Spec first. Goal, non-goals and milestones, each with a success criterion you can verify.
- Tests before features. Deterministic mock providers, so every AI project builds, tests and demos with zero API keys.
- Evals are a deliverable, with mock and live numbers labelled honestly.
- Agents get guardrails: tool allow-lists, step and cost budgets, and a human approval step for anything that writes.
- Observable from day one: structured logs, traces and metrics.
- Languages: TypeScript, JavaScript, Python, Kotlin, SQL
- Backend: Node.js, Express, Fastify, FastAPI, REST, webhooks, JWT / OAuth
- Frontend: React, Next.js (SSR / SSG / ISR), Vue, Tailwind CSS
- AI: LLM APIs and tool use, MCP servers and clients, RAG (pgvector, BM25, hybrid search, rerankers), LangGraph, evals
- Data: PostgreSQL, MongoDB, Redis, SQLite
- Cloud and DevOps: AWS (CDK, Lambda, API Gateway, IoT Core, ECS / Fargate, S3), Docker, GitHub Actions, Azure DevOps
- Quality: Vitest, Jest, Playwright, pytest, OpenTelemetry, Prometheus
The fastest way to reach me is email or LinkedIn.
Earlier projects (2020 to 2024)
- Medium Vanced: full-stack blogging platform (React, TypeScript, Express, PostgreSQL, Prisma)
- Paytm Wallet: digital wallet with transfers (React, Node.js, MongoDB, Docker)
- teckollab-transaction-dashboard: transaction data dashboard (TypeScript)
- Blockchain experiments: chain-query, sookshman, SignaKey, voting-dApp (Solana and Ethereum, Solidity). Certified Blockchain Developer, Kerala Blockchain Academy.