Team Lead, Stark Digital · Open to the right opportunity
I build thesystems that othersystems depend on.
I lead seven engineers at Stark Digital across three tracks: two seniors on a live healthcare platform, an AI engineer embedding intelligence into it, and four more turning out working prototypes most weeks. 6 years deep in the JavaScript ecosystem, with enough React Native to drop to native bridging when the JS layer runs out of room.
01 — Behaviour
The work is under NDA.
The behaviour is not.
Every system below is client-confidential, so there are no screenshots to show. These are the mechanics instead — which is the part worth showing anyway.
Public sector · statutory approvals
42 → 1
Forty-two independent departments, one consolidated certificate.
The hard problem was never the workflow — it was the 42 front doors. Every department had its own non-technical staff who had to adopt this mid-cycle against a deadline fixed in law, so it had to be obvious enough that nobody needed training.
Collecting consent from 42 independent government departments by hand took weeks and stalled part-way. The portal collected all 42 in parallel and issued one certificate within 48 hours.
Pattern · zero-login identification
Login, removed
Identity resolved at the point of contact, before the form opens.
Submission volume dies on account friction. The person filling this in has seconds and no patience, and every credential you ask for is a place they stop — so I took the account out of the path entirely. A tap on a physical token opens the form with the submitting organisation already resolved and scoped to what it is allowed to see. No sign-up, no password, nothing between someone and the thing they came to do.
Organisation resolved
Submitting the form previously required creating an account and signing in. Tapping a phone to a physical token now opens the form with the submitting organisation already resolved.
Pattern · multi-master sync
Any source, all in sync
Three third-party systems, every one writeable, none of them authoritative.
Each side of the workflow lives in a different system and none of them will give theirs up, so there is no source of truth to appoint. Every system is treated as a writeable master: a change originating anywhere is reconciled centrally and pushed out to the rest — and the write that lands there must not echo back and re-trigger the change it came from. Conflict resolution and echo suppression are the entire job.
02 — The practice
Five engineers. A Friday deadline.
I run the proof-of-concept pipeline at Stark Digital as a training programme. Engineers get a real problem, a contract-first architecture and a Friday deadline. I set the architecture and review every module; they build and own it. These are theirs — shipped, demoed and public.
Retrieving a specific fact from a large unstructured document repository is an operational bottleneck.
- Four-layer retrieval pipeline — vector similarity, metadata filtering and neural re-ranking — returning verified answers with exact page-level citations
- Row-level security for hard data isolation between tenants
- Redis semantic caching cuts LLM token spend by up to 50% and serves repeat queries in under 10ms
- Model-agnostic: swaps between OpenAI, Claude, Gemini and Llama 3 without code changes
Motion sensors switch the lights off on people who are sitting still. Static climate control wastes energy on empty rooms.
- Computer vision for continuous headcount, so stationary occupants are never missed
- Runs entirely on the local network — sub-100ms hardware response, no cloud dependency, no data leaving the building
- Staggered vacancy shutoffs and dynamic HVAC scaling, with live dollar-savings telemetry for facilities and ESG reporting
- Retrofits in parallel with existing switchboards — no rewiring
Constituency offices process dozens of walk-in grievances a day in Marathi, Hindi and English. Urgent cases get lost in the queue.
- Real-time speech transcription across three languages, plus OCR for handwritten letters
- Extracts a structured dossier per visitor — intent, sentiment, urgency, summary
- Live dashboard auto-sorted by urgency, with repeat visitors flagged for one-click routing
- Every AI component sits behind a swappable interface, so it runs on cloud models for demos and fully offline models in production — no rewrite
Matching paper receipts against bank statements by hand costs hours every week.
- Camera capture extracts vendor, date and total from any receipt
- Matching engine pairs each receipt to its bank transaction automatically
- Runs on-device — financial data never leaves the phone, and it works offline
Rage-bait and algorithmic manipulation are invisible while you are inside the scroll.
- Rolling 30-second sentiment analysis of on-screen content, entirely on-device via Gemini Nano and Apple Foundation Models
- Overlays flagged content and nudges a break rather than locking the phone
- Zero network permissions — it structurally cannot leak what it reads
- Auto-pauses on banking, password and secure apps
03 — Track record
Six years, three companies.
React Native depth throughout — including the parts where you have to leave JavaScript entirely.
Team Lead
Stark Digital Media Services, Pune
Leading architecture for a live healthcare platform and public-sector delivery. Seven direct reports across three tracks: two senior full-stack engineers on product delivery, one AI engineer on AI integrations, and four engineers on a proof-of-concept pipeline I run as a training programme, shipping a working prototype most weeks.
Lead Software Engineer
Kumowiz Technologies
Managed complete Android and iOS application release lifecycles. Spearheaded technical evaluation of critical analytics libraries before standardizing integration. Engineered and maintained React Native Bridges for high-performance audio/video processing.
Full Stack Developer
NO Brain Solution
Delivered high-fidelity interactive animations using HTML5 Canvas. Managed end-to-end delivery lifecycles from work distribution to UAT and Production deployments. Established coding standards and facilitated client communication.
Dgraph · SPRINT · SEB Cares
Earlier work
UI for cloud administration modules of an open-source graph database. Google auth integration, unit and integration test coverage, bug fixes in the internal Typhoon UI library. High-fidelity interactive animations in HTML5 Canvas and Anime.js, performance-tuned for fluid rendering across devices. Utility service apps and admin portal. Primary liaison between Operations and IT, owning delivery from requirements through UAT to production. B.Sc. Computer Science, Pune University.
04 — Contact