experience

The long version, reverse-chronological, with the footnotes and the side-quests that don't fit on a résumé.

Senior Product Engineer, Shoppin

Oct 2025 – now · Gurugram, India · full-time

came in to do

Own the systems behind Shoppin's AI fashion product as it grew from a discovery/try-on app into a full commerce and custom-manufacturing platform: the feed, the try-on, the pricing, the payments, and eventually the machine that turns a generated garment into a factory-ready sewing pattern. Roughly: everything between the content firehose and a real garment on someone's doorstep.

shipped

• Pattern-making engine · garment image + spec → fully-graded, production-ready ASTM sewing pattern (CAD-level DXF) in minutes; hybrid LLM-picks-operators + deterministic geometry, validated against an industry CAD oracle
• AI custom-clothing platform · image → flat-lay + title + deterministic pricing, plus 'imagine' conversational ordering (chat, reference images, try-on) → made-to-order
• Discovery & recommendation feed · 6-source blend over 250K+ items, Zilliz vector search, a two-stage Gemini query predictor, and a 10-stage content pipeline
• Commerce & payments stack · order lifecycle across Razorpay/PayPal/Stripe, native coupon engine, partial-COD with proportional refund-split, unified cross-gateway refunds
• Distributed scraping platform · 10+ microservices across 55+ fashion domains behind a 50M Bloom filter and an adaptive-concurrency Go browser pool
• Virtual try-on · 4-provider racing architecture (Google Cloud VTON, FashionAI, fal.ai, OpenAI) for 20K+ daily users; cut /profile/me from 6.8s to 500ms (14×)
• Eval harness for the AI surfaces · reference-set comparisons for image chats, the imagine flow, and pattern-engine output so quality never silently drifted

learned

• The hardest part of a recommendation system is deciding what "personalized" means for a brand-new user
• Provider racing beats picking the best provider; latency *is* the product
• Deterministic beats clever: LLM-only pricing gave the same tee two prices; a model-backed layer fixed it
• You don't automate the expert away; you hand them a near-perfect starting draft and let their judgment finish it
• Redis Lua is the cheapest distributed lock you'll ever write
• PG COPY protocol is 10–100× faster than executemany; more people should reach for it

fun things

• ~2,090 commits across 12 repos in ~9 months: Python backends, a TS API, two Next.js apps, a React Native app, and GPU/ML serving
• The pattern engine took a master pattern-maker from 2–4 drafts a day to 9–10
• Built a two-stage query predictor that guesses your next search before you type it

Founding AI Engineer, Verbaflo.AI

Dec 2024 – Sep 2025 · Gurugram, India · full-time

came in to do

Show up on day one to the brief "we need a backend" and leave with every system the product needs: architecture, schemas, real-time messaging, RAG, LLM serving, eval pipelines. The classic founding deal: own everything, hand it off as the team grows.

shipped

• Multi-modal agentic platform backend (Chat / Email / WhatsApp / Voice) · FastAPI, async throughout: core architecture, schemas, and real-time messaging, scaled to 20K+ concurrent sessions
• Hybrid RAG pipeline with Milvus · query expansion + reranking + parallel retrieval across multiple knowledge bases; sub-200ms retrieval at 95%+ context accuracy on multi-turn conversations
• Production LLM infrastructure · a multi-provider gRPC service (OpenAI, Anthropic, LLaMA, Qwen) with fault-tolerant fallbacks and per-provider rate-limit awareness, adopted as the standard across every team
• Real-time messaging layer · unifying four channels behind one assistant, with presence and delivery handling so a WhatsApp thread and a voice call feel like the same conversation
• Evaluation pipelines · regression detection on every model swap, so a provider or prompt change couldn't quietly make the bot dumber

learned

• Founding-eng is a speed game: write for rewrite; nothing you ship survives year two anyway
• Multi-provider LLM fallback isn't optional; models fail silently, and they fail often
• RAG quality is a query-expansion problem more than it is a retrieval problem
• One backend, four channels: the hard part is making them feel like one assistant, not four bots
• Eval pipelines kill 80% of the "why does the bot sound dumb today" debugging

fun things

• Scaled to 20K+ concurrent sessions on FastAPI before anyone thought to benchmark it
• Built the whole gRPC LLM layer on a weekend because the team kept hitting rate limits
• The LLM infra I shipped quietly became every team's default, the best kind of adoption

Founding AI Engineer, Sidecar

Jul 2025 – Oct 2025 · Remote · Contract · contract

came in to do

Build a freight-forwarding agent that could run a logistics desk on its own: read the emails, understand the ask, chase the missing details, get the quotes, and pick the best one, all without a human babysitting each step.

shipped

• 24/7 autonomous freight-forwarding agent · reads inbound emails, uses an LLM to understand intent, and runs structured extraction over messy PDFs, scanned rate sheets, packing lists and BOLs to pull shipment details
• Auto-reply + memory · chases missing info conversationally, with a long-term memory layer that remembers each client's preferences across shipments
• Agent harness from scratch · orchestration, tool use, retries, and guardrails
• Agentic quoting loop · fans out to partner freight-forwarder agents, ranks quotes on price, transit time and reliability, and returns the best option with no human in the loop

learned

• Real-world documents are hostile; a scanned rate sheet will break any extraction that assumes clean input
• An agent that says "I need one more thing" beats one that guesses and gets it wrong

fun things

• Turnaround went from hours to minutes and drove 3× outbound volume

Founding Engineer, PrudentBit

Jun 2023 – Aug 2024 · Hybrid · full-time

came in to do

First real founding gig: show up, learn how to ship enterprise software, and convince Toyota and UCO Bank that this is something they can run on.

shipped

• Secure file-sharing platform on Azure · 6+ microservices with enterprise security: SAML SSO, OAuth, 2FA, DLP, malware scanning
• Backend rewrite · 130+ REST APIs, Celery/RabbitMQ work queues, chunked uploads, full-text search, and a RAG-powered document interface
• Performance · cut API latency from 1s to 30ms and page loads from 2s to 0.4s on a backend that couldn't scale before
• Enterprise deployments · Toyota and UCO Bank, with the security surface (SSO, DLP, malware scanning) that those deals actually turn on

learned

• Enterprise clients tell you the real requirements six months in, so start with the flexible version
• A 33× latency win (1s → 30ms) is usually hiding in one query nobody dared to rewrite
• The security work nobody wants to own (SAML, DLP, malware scanning) is what unlocks the enterprise deal
• Azure is not AWS; both are fine, neither is magic

fun things

• First time selling into Toyota; still fascinated by the procurement processes of companies older than me
• Built full-text search from scratch because Elasticsearch felt like overkill for our corpus

contract & freelance

Selected contract work for founders and teams who needed a system built fast, all shipped, all running in production.

• Convin.AI · 25% query-performance improvement on live dashboards
  Real-time CDC pipelines feeding enterprise contact-center analytics: terabytes of Postgres data, live-streamed into a warehouse built for fast aggregation.
  • Debezium + Kafka Connect captures Postgres row-level changes and streams them to Kafka with at-least-once delivery
  • Go consumer workers transform and batch-write to Apache Pinot, with idempotent keys so replays don't double-count
  • Schema-evolution handling that auto-adds Pinot columns as upstream tables grow, so ops doesn't ship a migration per column
  • Backfilled the historical data without downtime using a blue/green table switch
  Kafka Debezium Go Apache Pinot PostgreSQL
• Event-management & wedding platform
  Built the backend for an event-management / wedding platform (including a full job portal), plus LMS-style features for their programs.
  • Job-portal backend from scratch: listings, applications, and role-based access across candidates, organizers, and admins
  • LMS-style features: a quiz engine with timed sessions and scoring, plus an assignment-submission pipeline
  • Secure HLS video delivery: m3u8 playlists served from signed S3 URLs with short-lived tokens and adaptive bitrate
  Node.js Express HLS streaming PostgreSQL

Prefer the short version? /about.md →