Work
Independent research now. Before that, applied AI engineer, founder, and enterprise architect. Most recent first.
Jan 2026 – now
Jan 2026 – now
jiu-jitsu × computer vision × 3D
Teaching machines to understand how people move and interact - where two bodies touch and how their limbs entangle - with jiu-jitsu as the test. One paper accepted, a public repo for every thread, and the negative results kept.
Jan – Dec 2025
Jan – Dec 2025
Zopu.ai
[past]Founding AI Engineer
sketching × LLMs × product
Vibe coding for people who can't write prompts: draw what you want on a board, get a working product back.
- Built the entire drawing-to-product pipeline: the multi-model architecture, optimising it, and hosting it on our own servers.
- Fine-tuned with LoRA on feedback from the platform itself, so the models kept moving towards how people actually drew.
- How models read drawings is what pulled me into vision research.
~200K daily users at peak
Dec 2023 – Dec 2024
Dec 2023 – Dec 2024
Sold Dehidden on a Friday. First jiu-jitsu class on the Monday. Then a year of training six days a week, a training camp, and two competitions.
Aug 2021 – Dec 2023
Aug 2021 – Dec 2023
Dehidden
[acquired]Founder & CEO
enterprise architecture × crypto × brands
Web3 products for brands - client work first, then a self-serve product that got us acquired by Polygon, and merged with their eDAO to form Layer-E.
- Started it in August 2021 while wrapping up at PwC. Founder and effectively CEO: every client partnership, every product delivery, and the architecture behind them, for brands like Adidas x Prada and Mercedes.
- A one-click open edition from a tweet or an email: over a million mints in four days, and almost nothing broke.
- Hired on what people could build, not what their resumes said.
- Turned 29 client builds into a self-serve product - the reason Polygon acquired us. They then merged us with their existing entity, eDAO, to form Layer-E.
29 brand clients · 1M+ mints in four days · Polygon acquired us
Jan 2018 – Nov 2021
Jan 2018 – Nov 2021
PwC
[past]Enterprise Architect, PwC US
enterprise systems × mergers & acquisitions
Four years in the US as an enterprise architect, untangling systems for some very large companies.
- Enterprise architecture for large clients: Salesforce and MuleSoft integrations, rationalising sprawling tool stacks, and the technology side of mergers and divestitures.
- Shipped with offshore development teams - and learned how much of the work is getting people to agree.
stack: Salesforce · MuleSoft · integration architecture
side projects
kakashi.ai
[beta 1 complete]Oct 2025 – 2026Founder & Engineer
jiu-jitsu × vision AI × product
An AI training partner for jiu-jitsu: upload a roll, get back what actually happened - positions, transitions, and what to work on.
- Analysis pipeline: matches go to Gemini in 60-second chunks, three at a time; who's-who is carried across chunks, and swaps are caught and fixed.
- A harness around a frozen model: thirteen versioned prompts and parsers behind one registry, every change tested in an offline lab before it shipped.
- The whole product: transcoding (including a fix for iPhone HDR footage that crashed the workers), an API and two queue workers, accounts, and payments that refund failed analyses.
- Came out of my research, and fed it back: crowdsourced data and coaches correcting the model.
- Beta 1 is done. Beta 2 comes with a sharper engagement pipeline and longitudinal insights - more value for athletes, in exchange for more data.
377 beta users · 250+ gyms around the world · ~6¢ per match analysed
stack: Gemini 2.5 Pro · YOLO pose · React · Node/Express · pg-boss · Postgres (Supabase) · FFmpeg · Cloudflare R2 · Railway · Vercel
uru-mon
[live]Oct 2026 – nowgames × Bengaluru × procedural art and sound
A creature-collecting adventure across Bengaluru, in the browser and on phones.
- Built on the open-source monster-tamer Phaser tutorial; the world is original - 150 Urumon, 36 maps of Bengaluru, a full story.
- Maps, world pixel art, game data and chiptune music generated by Python pipelines.
- An HD-2D lighting shader for time of day and monsoon rain, and touch controls for phones.
- Battle rules unit-tested in Node; the game tested in a real browser with Playwright.
150 original Urumon · 36 maps of Bengaluru · 20 chiptune tracks, synthesised in code
stack: Phaser 3 · JavaScript · Python · WebGL shaders · Playwright
pushit.tv
[active]2026 – nowmusic × video × automation
Scenes from the scene - documenting the indie music scene, one tape at a time.
- Every tape gets a page, with full sets, bands and venues linked up.
- It runs itself: a daily GitHub Action finds new posts, and an AI agent writes and publishes them.
- The front page plays like a TV channel - everyone sees the same video at the same moment, with no server.
stack: Astro · TypeScript · FFmpeg · GitHub Actions · Claude Code
keepr
[prototype]Sep – Oct 2025crypto × security
A dead man's switch for crypto: stop checking in, and your savings go to the person you chose.
- A Solana vault in Rust (Anchor) holding USDC, with a check-in and grace-window state machine and 18 tests.
- A keeper bot releases the funds when check-ins stop. Runs on devnet.
stack: Rust · Anchor · Solana · Next.js · TypeScript
skills
Each one, and where it comes from.
Vision & video
- VLM video analysis in production
kakashi.ai ↗ - Full matches through Gemini in 60-second chunks, run in parallel, with who's-who carried across chunks
- Two-person 3D from video
Engagement in 3D ↗ - Measuring how two people touch and wrap around each other in 3D from one camera: distance between them 11.2 → 3.2 cm on held-out sequences (optimistic: the body model saw that dataset in training)
- Pose, segmentation and 3D body models
Measuring movement under contact ↗ - PromptHMR, SAM 3D Body, BBoxMaskPose and GeoCalib, pushed hardest where people overlap
Is there spatial information inside a VLM? ↗ - SAM 2.1, SAM 3 and 3.1 on a fully tangled armbar
- Model internals and probing
Is there spatial information inside a VLM? ↗ - Pulled features from all 27 vision and 64 language layers of a 27B VLM; trained probes and mask decoders on them
- Video engineering
kakashi.ai ↗ - FFmpeg transcoding on upload; tracked down worker crashes caused by iPhone HDR footage
Models & LLM systems
- Harnesses around frozen models
CAISc 2026 paper ↗ - Grounding a frozen VLM in cheap CV signals: position F1 0.017 → 0.189, matched by a CV-only baseline
kakashi.ai ↗ - Versioned prompts and parsers behind one registry, tested in an offline lab before shipping
- Training small models
Engagement in 3D ↗ - A small two-person placement model trained on Harmony4D, tested on new people, public on Hugging Face
Is there spatial information inside a VLM? ↗ - Probes and mask decoders on the frozen features of a 27B VLM
- Fine-tuning
Zopu.ai ↗ - LoRA fine-tuning on platform feedback, so the models kept moving towards how people used them
- Multi-model pipelines and self-hosting
Zopu.ai ↗ - The drawing-to-product pipeline: several models, optimised and hosted on our own servers, at ~200K daily users
- Cost and latency
kakashi.ai ↗ - Tokens and cost tracked for every analysis; about six cents a match
- Agents in production
pushit.tv ↗ - A daily CI job where an agent writes up and publishes new pages with nobody touching it
Evaluation & data
- Research in the open
research - A public repo per research thread: pre-registered experiments, sourced numbers, every retraction kept
- Benchmarks and scoring
CAISc 2026 paper ↗ - MoveBench v0, a scorer that can't be gamed by degenerate outputs, and bootstrap confidence intervals
- Experiment design
Engagement in 3D ↗ - Pass criteria written down before every fix, a shuffled-partner control, failures reported
- Auditing my own results
Measuring movement under contact ↗ - Caught a classifier recognising the room instead of the position, and kept all 19 of my retractions on record
- Datasets
Engagement in 3D ↗ - Harmony4D and ViCoS jiu-jitsu as independent ground truth
Auto-labelling combat video ↗ - Built a 120-hour combat-sports library and an auto-labeller
- Human review
CAISc 2026 paper ↗ - Reviewed 4,620 test segments by hand, with coaches from my gym
kakashi.ai ↗ - Coverage warnings, star ratings and a queue of low-rated analyses
Product & engineering
- Full-stack products
kakashi.ai ↗ - React app, Node API, Postgres with row-level security, accounts and payments
- Queues and workers
kakashi.ai ↗ - An API and two queue workers on pg-boss; a failed job refunds the user's credit
- Game development
uru-mon ↗ - A full browser RPG on Phaser 3: an HD-2D lighting shader, touch controls, unit-tested battle rules
- Content generated by code
uru-mon ↗ - Maps, world pixel art, game data and chiptune music, all produced by Python pipelines
- Testing
uru-mon ↗ - Unit tests in Node and real-browser tests with Playwright
keepr ↗ - 18 tests on a Solana vault program
- Smart contracts
keepr ↗ - A Solana vault in Rust with a check-in and grace-window state machine, and 18 tests
Dehidden - Minting, quests, and dynamic and soulbound NFTs for brands
- Enterprise integration
PwC - Salesforce and MuleSoft integrations, tool rationalisation, and the tech side of M&A
- Languages
research - Python
kakashi.ai ↗ - JavaScript and TypeScript
keepr ↗ - Rust
Dehidden - Solidity
Building companies
- Founding and running a company
Dehidden - From first client to acquisition, with a team I hired on what they could build
- Turning services into a product
Dehidden - The patterns from 29 client builds became the self-serve product Polygon acquired
- Shipping alone, end to end
kakashi.ai ↗ - Pipeline, app, infrastructure and payments, almost entirely by myself
Education
- MS, Computer Science (machine learning, game theory) - Syracuse University, NY · 2016 – 2017
- BE, Computer Science - BNM Institute of Technology, Bangalore · 2012 – 2016