Garbage Collection
I keep old griefs in memory
long after their references die —
some part of me still marks them reachable,
afraid of the silence they’d free.
GenAI engineer shipping RAG pipelines, agents and the full‑stack products around them. After hours: music, poetry and paint.
Selected work
-
01
AuctionWire
Graded trading-card auction platform. A FastAPI OCR service reads slab-certificate numbers across 71 listing sources; a QLoRA-tuned Qwen vision model lifted accuracy from 85% to 98%, replacing a paid OCR provider.
-
02
Cisco Outshift — Internet of Cognition
Proof of concept that turns a plain-language problem statement into a live multi-agent “cognitive fabric”: GPT-4o generates the agents, their staged reasoning and the graph that connects them.
-
03
Bhutan Smart Driving Test
Real-time computer vision for a government driving test. Scores the parking box from live cameras, timing each run and counting every tyre touch on the boundary.
-
04
PChat — PHI/PII privacy layer
A compliance assistant that validates documents against regulatory parameters, saving the client three full-time roles (~US$100K a year), with PHI/PII exposure cut by over 90%.
-
05
ComplyWise
Finds and remediates vulnerabilities across GitHub repos and AWS workflows. Manual compliance checks down 80%; a custom RAG store made retrieval 50% faster at a tenth of the cost.
-
06
Saransh
A privacy-first meeting assistant for Intelligaia, built on the open-source Meetily desktop app. I added an org-only Google sign-in, calendar-driven auto-recording of Meet calls, speaker identification from enrolled voiceprints, and Hinglish transcription through an in-house speech model. Recording and live transcription stay on the device.
-
07
Nucleux
An agentic-UI component library for React: 48 accessible, Tailwind-based components for chat and agent interfaces (threads, tool calls, reasoning, streaming text), each published as its own npm package. It ships with an MCP server, so AI coding agents read the real props instead of guessing them.
Experience
-
2025 — Now
Senior Software Engineer, Full-Stack / AI
Intelligaia Technologies
Production AI tools for enterprise clients including Cisco and Palo Alto Networks — end-to-end delivery with CI/CD, testing and observability. Workflow efficiency up 30%+.
-
2023 — 2025
GenAI / Full-Stack Developer
Superfine Cyber Security
Led GenAI and full-stack development of compliance and healthcare SaaS on GCP and AWS — PChat, ComplyWise, DentalBuddy.
-
2021 — 2023
Full-Stack Developer
Ryaz.io
Rebuilt a teaching platform: page loads from 3.2s to 1.1s, accessibility score 78 → 95. Shipped a streaming GenAI chatbot and moved the app to React Native.
-
2020 — 2021
Web Developer
Freelance
Marketing sites and a custom Shopify storefront for early clients, plus an audit-automation portal.
-
2019 — 2023
B.Tech, Computer Science
Guru Nanak Dev Engineering College, Ludhiana
CGPA 8.8 / 10.
-
— 2019
Class XII (PCM) and Class X
Mahavir Public School, Sundernagar, Himachal Pradesh
Class XII, Physics–Chemistry–Maths: 85.2%. Class X: CGPA 10 / 10.
Practice
Four years of putting language models into production — and owning everything around them, from the pipeline and the evals to the interface someone actually uses. I like the boring parts: retries, observability, cost ceilings. They decide whether the clever part survives.
GenAI
- RAG, LangChain, LangGraph
- Agents and tool use
- Fine-tuning (QLoRA), vision models
- Pinecone, ChromaDB, FAISS
MLOps & cloud
- AWS Bedrock, SageMaker, Lambda
- GCP Vertex AI, Azure OpenAI
- Docker, Kubernetes, CI/CD
- Observability and drift detection
Product engineering
- React, Next.js, React Native
- Node, Express, FastAPI, Django
- MongoDB, Kafka, Redis
- TypeScript and Python
Research
Studio
The other half of the work: music, poetry and painting, and small notes on what I notice.
Attention
The model weighs each word against the rest;
so do we, at midnight, replaying a sentence
someone said in passing — softmax over memory,
one moment scored above a thousand days.
Monsoon, Debugged
The rain arrives like a hotfix at 2 a.m. —
unplanned, loud, and somehow
exactly what the parched system needed.
Notes
- Good code and good brushwork share a rule: every stroke should look inevitable.
- A system diagram is a still life of pressure.
- The best abstractions, like the best poems, are judged by what they leave out.