Open to backend and AI engineering roles, remote
Backend engineer. I build the parts of a product users never see and always feel: search infrastructure, data pipelines, and LLM systems that hold up under real traffic.
About 4.5 years across Wayground (Quizizz), Apollo.io and Walmart, mostly in Go and Python. Recently I've been going deep on agentic systems, both at work and in my own projects. When I'm away from a keyboard I'm usually on a trail, under water, or on a motorbike.
Moved a critical, unowned indexing pipeline into an owned, observable service and re-indexed 120M records with no maintenance window. Reads go through an alias, so cutover and rollback are each a single atomic swap.
120M records0 downtime1-step rollback
Go · MongoDB CDC · Airflow · S3 · SQS · OpenSearch
Led a cross-team build that opened the platform to ChatGPT, Claude and Superhuman over MCP. Encrypted opaque tokens in Redis instead of JWTs so access can be revoked instantly, with PKCE required for every client. Once registration existed, the Claude integration took under a week.
3 integrations~8 weeks0 auth regressions
Go · Redis · OAuth 2.0 · PKCE · MCP
An agentic pipeline that verifies CRM contacts and reads public school-district documents (board minutes, improvement plans) with OCR and LLM extraction. It produces structured signals about curriculum needs, competitors and decision makers, and a dashboard the sales team used live on calls.
45K / 70K contacts verified3 deals in month one
Python · LLM extraction · OCR · Playwright · Qdrant · BigQuery
Took an ML proof of concept to production as a pipeline where every stage retries independently over SQS. Cut LLM spend with model tiering and by replacing some LLM calls with rules. Separately, reworked a core API with cursor pagination to bring P95 latency down 70%.
−70% P9525 US states live
Python · Go · SQS · MongoDB
Built one of Apollo's most-requested features end to end: resolving anonymous website-visitor IPs to companies, aggregating intent signals, and indexing them beside a 256M-record search index without touching it. Taken from proof of concept to production.
1M+ req / week~20% → 40–50% match rate
Ruby on Rails · MongoDB · Elasticsearch · Sidekiq · Snowflake
Things I've built for myself this year, mostly to see how far agents can go when they're built carefully.
An AI trip planner where every factual claim carries its source page and the date it was read. Visa rules come only from official government sites; routes and stops are researched live and cached.
TypeScript · Next.js · web-research agent
Harvests roles, ranks them against my profile, and spends paid contact lookups only on the best leads. It drafts emails into a queue I approve by hand, because nothing sends without me.
Python · SQLite · MCP · Gmail API
Books appointments over a phone call, streams a live monitoring dashboard, lets a human take over mid-call, and warm-transfers with a spoken summary.
LiveKit · Python · Deepgram · Cartesia · Next.js
A hotel sales agent that qualifies an enquiry, recommends a programme and closes on a line-item proposal. Each property's behaviour lives in a config you edit in the browser.
Node · tool-calling LLMs · OpenRouter
A workflow–agent–tool pipeline that turns a topic into a captioned short video.
Python · workflow-agent-tool pipeline
Competitive programming on the side: CodeChef 5★ (max 2090), Codeforces Specialist, ICPC 2020 Asia-Amritapuri regionalist.
I'm happiest learning something by doing it badly at first. A few of those things:
If you're building something with hard data, search or agent problems in it, I'd like to hear about it.
Bengaluru, India (IST). Comfortable overlapping with US and EU hours. I usually reply within a day.