Work

Tools I use myself, companies I started, and places I've worked. Repos are public unless marked private.¹

Blindfold. With Arnesh Banerjee (his repo, private). Launch video. "An agent you can point at your bank, your inbox and your health records, because nothing on the screen ever leaves the machine." A browser agent that redacts pages on device and sends only stand-in values through one audited exit. Co-led a six-person SIH 2026 team (ISRO problem 26171); most commits and merged PRs in a ten-day build. Owned perception, the agent runtime and measurement. Private.²

maiku. One runtime for Python functions, local LLMs, PyTorch modules, MCP servers, A2A agents, CLIs and people, through one streaming event contract. Capability grants can only narrow. Its real examples run Gemma 4 on llama.cpp, a CNN trained in the repo on MNIST to 98.2%, and Anthropic's MCP filesystem server. Private.³

ChromeClaw. A browser agent: one schema-validated action per step, a policy layer that blocks or escalates credential, payment, messaging and destructive actions, and a JSONL trace of every step.⁴

Horizon. A local decision workstation for the next 12 months: every opportunity with its evidence, Monte Carlo over an action graph, strategy search, and a replay of any simulated world as a video. Private.

Entity resolution, Amazon ML Challenge 2026. Match 2.2M businesses to 10.3M records across the US, India and France: fine-tuned multilingual bi-encoder blocking, two cross-encoders, a two-stage stacker, one-to-one assignment. On an H100.⁵

CACCS-AI. With Suramyaa Sarkar (University of Oklahoma): narratives into causal loop diagrams, system archetypes, game-theoretic equilibria and simulations. React, D3, FastAPI.

A digital twin. A private chat with a persona of myself, built from my own archive; Supabase, persistent memory.

batcave. A private Mac app: monthly goals and work sessions, food, macros and workouts, and world news from three outlets. Tauri, Rust; ships as a `.dmg`.

Logbook. A personal telemetry app for the Mac: work, calories, protein, caffeine, sleep, with raw timestamps and no cloud.

sbcmd. A local-first Windows command center with a modular part registry and local AI providers.

ReasonTrace. A visual debugger that turns agent logs into a seven-event graph, with heuristic checks for loops, unplanned actions and weakly grounded answers.

Casca. A desktop agent runner driven by screenshots, with typed actions and dry runs.

Sefai. Offline GGUF inference in about 150 lines of Rust, with per-layer GPU offload and CPU, CUDA, Vulkan and Metal builds.⁶

Stockfih. Stockfish 18 in the browser (WASM), depth 8 to 14, with engine facts kept separate from the LLM's commentary, explained three ways (plain, "like I'm 1200", tactical).

CuteTuner. A local-first vocal tuning studio. Nothing leaves the Mac. Learns your taste from A/B ratings after 20 comparisons.⁷

ParaReport. Kolkata civic reports in Bengali, Hindi or English, turned into receipts and complaint packets, with local Gemma and signed federation between instances. Gemma hackathon, July 2026.⁸

lifesim. A Monte Carlo life simulator that refuses to predict: 200 to 6,000 lives per run, and the policy you'd regret least in the worst case.⁹

OpenMind. An AI code editor with the model inside it. I was CEO, with two co-founders. A Django version was live in under a week on ₹5,000. We stopped after three months: real-time sync and fast inference were more than we could build yet.¹⁰

faculza. (2023) A website builder you talk to: describe a component, it appears on a drag-and-drop canvas, then publish or export the code. I built it alone, with a co-founder from a community of founders with ₹10–100+ crore outcomes. An angel investor booked a meeting and didn't come. We split over direction.¹¹

NashBase. (Jan to early Apr 2026) With a potential co-founder. We called it "a research lab first, start up second." The thesis was a four-way labor market: humans and AI agents on both sides, each posting work and each doing it, in one network. David, our AI agent, matches every task to the best executor, human or AI, then runs the whole job: scope, milestones, coordination, delivery. "Think of him as the operating system for your work."

The bet underneath was data. Capture the full path from an idea to money, with money as the forcing function that filters out noise, and train David on it, aiming at the world's best superintelligence. As open models and compute get cheaper, that data is the moat.

What shipped, in order: AI talent matching (10 Mar); the unified market and David (14 Mar), with a demo that returns the top five freelancers for a plain-English request in under 10 seconds; an AI project manager (15 Mar); a move to a Django and Supabase backend (30 Mar); job matching from a parsed resume, and a Resume Roaster that roasts your CV and ends with one real piece of advice (2 Apr). Pitched AJVC and Antler, met Next Bharat Ventures, and got a referral to F Inc from its founder by cold DM. Then we stopped.¹²

A research collective. (wound down) Founded a nonprofit AI research collective with people from Sarvam AI, AI4Bharat, ISI and IIEST in April 2026. It didn't last.

A college AI research group. (Feb 2024) For AGI-level research, from a college. 150 members in under two weeks; about 125 are still in it.

Earlier attempts. (2019 to 2021) A Shopify dropshipping store at 14 (no sales). An app for remote school at 15 that my class teacher committed to adopting; we couldn't build it yet. Both on History.

Jadavpur University. Research collaborator, Aug 2026 onward. Indic LLMs and MT with Prof. Sudip Naskar's group; co-authoring a paper.

Unify (YC W23). Design partner, Aug 2026. Tested a pre-release agent platform and reported reproducible gaps in OAuth permission visibility, learned-memory review, action audit trails and per-action cost attribution.¹³

Heritage Institute of Technology. Undergraduate researcher, humanoid robotics, Jul to Oct 2025. Wrote and defended a proposal for a low-budget assistive humanoid robot through several faculty review rounds.¹⁴

buildspace. Season 5, Jun to Jul 2024. Built an offline LLM for Android that acts (messages, calls, music) solo in about three weeks. Left at the halfway mark when the program moved to marketing.