APM who builds · Stackbox
I build AI and analytics products for supply chains, and the tooling that tells us they actually work.
Now · Oct 2024 — present
Associate Product Manager on the analytics and control-tower platform used by leading FMCG and logistics companies across India and South-East Asia. I own the roadmap and the requirements, and I get into the technical details when that's what it takes to ship.
01 · AI quality
Every release of our analytics chatbot was a leap of faith. Nobody could say whether a prompt tweak or a model swap made answers better or quietly broke them. I designed and built an evaluation harness so that each release gets checked against the last one.
Every question, side by side. Latency, tokens for each pipeline stage and the generated SQL for both versions, plus one-click AI insights on what changed. The whole release at a glance. Token usage per question, where tokens go across the pipeline, and the % change for every question in the suite.
02 · Platform ownership
An enterprise analytics and control-tower platform for 5+ clients. I own the roadmap and the technical requirements across three capabilities: self-serve dashboards, an LLM chatbot for conversational deep-dives, and a real-time temporal alerting system that surfaces what needs attention before anyone asks.
Why did OTIF dip on Thursday?
3 North lanes missed dispatch cut-off after a vehicle shortage at DC-Ludhiana. SQL
03 · Product redesign · Setu
Setu is Stackbox's integration and data orchestration platform. Its promise is integrate once, connect many: ERPs, CRMs, e-commerce and legacy systems plug in once, and Setu connects, transforms and orchestrates their data into WMS, HR, planning and third-party tools. I led its end-to-end redesign, turning workflows that only engineers could operate into a product that ops teams can run themselves.
Setu in production
Enterprises across FMCG, retail, manufacturing and logistics run their data through Setu.
04 · Client integration · Philippines
I led the client-facing API integration for a major supply-chain company in the Philippines. It covered Master Data for 5 entities with full CRUD, plus the Operations endpoints. I also designed an automated email notification system that matched the client's exact encoding spec, and ran daily syncs with their team through go-live.
05 · Internal AI
Our most complex internal tools lived in a few people's heads. I distilled that institutional knowledge into domain-grounded, system-prompted Claude assistants. Now anyone can ask instead of waiting for the one person who knows.
A client's order batch failed ingestion overnight. How do I reprocess it?
Check the failed episode first. It's usually a schema drift in the SKU column.
1. Open the run and filter by status = failed
2. Fix the mapping in the transform step
3. Re-trigger just that batch, not the full routine
> cd ~/origins
I've always built things. Before product, it was GenAI data tools, blockchain cryptography, Android apps, cloud migrations, and an ML paper published by IEEE. That's why I can sit with engineers and argue about the right join.
Jan — Jun 2024
Built LangChain + OpenAPI tools for analysing Snowflake data, with interpretations and visualisations served through a web interface. Delivered EDA reports and predictive models to 4 clients.
# ask the warehouse in plain English $ chain = SQLDatabaseChain(llm, db="snowflake://sales") $ chain.run("Which region grew fastest in Q1?") → South, +14.2% QoQ · chart rendered
May — Jul 2023
Developed a proprietary email-ID hashing algorithm for accessing the Wayvy blockchain, and integrated ECC for secure, efficient user key generation.
$ key = ecc.generate(curve="secp256k1") $ wallet_id = hash("user@mail.com", salt) → 0x9f3a…c21e ✓ verified
Apr — Jun 2023
Shipped auto-population of multiple insurance forms from one consolidated user profile.
May — Aug 2022
Proposed cost-saving measures for migrating websites to AWS.
Published research · IEEE ESCI 2024
I co-authored "AI LipReader: Transcribing Speech from Lip Movements", presented at the 2024 International Conference on Emerging Smart Computing and Informatics (ESCI) in Pune and published by IEEE. The model uses CNNs for spatial features, LSTMs for temporal dependencies, and an adaptive attention mechanism. It was trained on MIRACL-VC1 and extended to LRS2 for a broader vocabulary.
Project
Trained models on a car-image dataset and built a web app around them. You upload photos of a car from two angles, and it predicts the damaged parts and estimates the repair cost.
# damage-assess · YOLOv5 $ detect("front.jpg", "side.jpg") → bumper, left fender · est. ₹18,400
Roots