History

Survey sites. I tried to earn my own money on online survey sites. I made close to nothing.

Binary options. In Class 6 I found binary options on IQ Option. That pulled me into equities, options and futures: price action, risk and reward, trading on margin, position sizing, expected value.

First code. QBasic at school, in BlueJ. I taught myself Java the same year and kept at it through Class 12: object-oriented design, data structures, algorithms.

Dropshipping. I studied sourcing, paid ads, funnels and unit economics in communities where operators made ₹10–50 lakh a year, then ran my own Shopify store. No sales. I lost the subscription fees and learned what an online business actually costs.

A trading system. U.S. equities on NYSE price history. First a rule-based ensemble: moving averages, RSI and volatility bands voting together. Then an LSTM. Backtests only, no real money; it beat the baseline strategies in simulation.

My first startup. In lockdown we wrote tests on Zoom, made a PDF, switched apps and uploaded to Google Classroom, and people missed deadlines after finishing on time. I wanted one app for classes, assignments and submissions. I put a team together and pitched it to my class. We couldn't build it yet.

School. National Gems Higher Secondary School, Kolkata, from Lower Nursery to Class 12.

A multimodal assistant. In the break before college: text, vision and voice through Hugging Face APIs, around Falcon, routed in Python on my laptop. The demo is on LinkedIn.

faculza. I met a co-founder in a community of founders with ₹10–100+ crore outcomes. We went through dozens of ideas and landed on prompt-to-feature. I built it alone: a canvas website builder with a chat sidebar. Describe a header and it appears; drag it, publish the page, or export the code. An angel investor confirmed interest and booked a meeting, then didn't come. We split over direction, not over the tech.

College. B.Tech in Computer Science and Engineering at Heritage Institute of Technology.

On GitHub. Opened the GitHub account that now holds 27 repositories.

SIH 2023. My first month of college. A disaster-response platform; I built the intelligence. DBSCAN finds clusters of distress alerts from social media, news and users, flags a likely disaster, and the map routes the nearest ambulances, police and relief units to it. A Falcon assistant answers questions during the incident. Shipped as an offline-capable PWA with role portals and per-incident chat rooms, in about 24 hours.

NASA Space Apps. My part of the team's project: a Unity game where you walk an astronaut across the moon, past craters, a rover and the lander, with physics and collisions. About 48 hours.

OpenMind. An AI code editor with the model inside it, built with two co-founders. I was CEO. We hadn't heard of Cursor. Before building, I sent one WhatsApp message to a few college groups. We put ₹5,000 into a domain, email and hosting (₹4,000 of it mine) and had a Django version live on Hostinger in under a week. I reached a senior Antler India partner on LinkedIn, who referred me to Antler India's leadership. We stopped after three months: real-time sync and fast inference were more than we could build yet.

An AI research group. A college group for AGI-level research. To join, you had to agree that AI is a real form of intelligence and could become conscious. It ran on WhatsApp and is quiet now.

Every market on Kaggle. NYSE, NSE and BSE stocks, options and futures, Bitcoin, crypto options, meme coins: gigabytes of history, with volatility and distributions at every window from 10 minutes to a year. The bot built on it beat buy-and-hold and naive momentum by multiples in backtests, with controlled drawdowns, and was designed to trade live through the Zerodha API.

Three papers. A new MCMC sampler made by composing three algorithms into one: better exploration per step than standard samplers, much slower to run. A fee-less blockchain whose miners are AI agents, with incentives in the protocol and human costs covered by minting, not fees. Federated learning where proof-of-work verifies every gradient update, so compute spent is training done.

buildspace S5. An offline LLM on Android that acts: it sends messages, places calls, plays music, with no cloud. I built it alone. When the program moved to marketing at the halfway mark, I left. I'd come to build.

SIH 2024. Top of the college round's pitch stage, with a working deepfake image detector behind the slides.

Shor's algorithm with light. A 4-qubit photonic quantum computer for factoring, worked out on paper from first principles: photon sources, encoding, gates, interference, measurement. A physics professor checked it and confirmed it works in ideal conditions. He kept the paper.

Debate of Thought. Several LLM agents each solve the problem, then critique and revise each other's answers for n rounds before the best one is chosen.

NSHM Innovathon. I built the AI advisor of a university search platform: tell it your scores, budget, location and priorities by text or voice, and it recommends universities from a QS-ranked database. About 18 hours.

MLX 2024. An inter-college Kaggle competition, one classification and one regression track. I led both for long stretches and finished 8th.

Synthesis of Thought. A controller decides how many agents a task needs and gives each a different way of thinking. Instead of picking one answer, it merges the strongest parts of each, crediting which agent contributed what.

MLX 2.0. Same competition, second year.

An answer engine. Like Perplexity: about 50 pages retrieved per query, a LLaMA model writes one answer with inline citations. Python, Streamlit, Groq, Bing Search. Being rebuilt with a TypeScript front end.

English to SQL. Talk to a live MySQL database in plain English: "create a table with name, age and city", "add James, 20, Bangalore". It writes the SQL, runs it and shows the table. My SQL professor checked it.

Fine-tuning without code. Upload model weights, a dataset and a run command; download the tuned model. TensorFlow and Streamlit.

A music critic. A national hackathon where investors were scouting teams for up to ₹1 crore. We cleared the pitch round; when the team couldn't build the core, I built it alone: any song in, pitch, overtones, rhythm and harmony out, then a report with strengths, fixes, similar songs, a score out of 10 and a forecast of whether it would be a hit.

Smart Bengal Hackathon. Through screening, then a prototype round to the final. I led the AI of a speaking coach: MediaPipe classifies posture frame by frame, the audio track measures pace, pauses and fillers, and an LLM combines them into live feedback. About 48 hours.

ISRO Bharatiya Antariksh Hackathon. Flood and drought spread from ISRO satellite time series: DBSCAN finds the emerging risk zones, LSTMs learn how they grow and shrink, and the map animates the past into a forecast. We didn't reach the final.

NeuroSymbolic Chain of Thought. Every reasoning step is checked with symbolic logic as soon as it is written. If a step fails, the error goes back to the model and only that step is regenerated, so one early mistake can't poison the whole chain.

A humanoid robot. A proposal for a low-budget assistive humanoid: joints, load-bearing structure, actuators, 3D design, control. I defended it through several rounds of faculty review.

SIH 2025. I led the team. We won the live pitch round in front of founders of million-dollar businesses with an AI platform for city traffic and operations.

Hack Heritage 2025. I led the team, and when it couldn't build, I built the whole thing: a map you talk to ("show future coral reef expansion here for five years") over endangered species, reefs and pollution, with past and future simulations, a carbon-credit marketplace, two-way conservation fundraising, citizen photo reports and an AI tutor.

A reasoning architecture. Started a paper on hierarchical reasoning models over RWKV-style backbones: planning, abstraction and tokens on separate levels, without quadratic attention.

Wrote, sang and produced a rap song, and released it. Another song followed in March.

NashBase, with a potential co-founder: a four-way labor market where humans and AI agents both post work and both do it, run by our AI agent, David, who matches each task to the best human or AI and manages it to delivery. The bet was the data: the whole path from an idea to money, to train David on. It shipped talent matching, an AI project manager, job matching and a Resume Roaster. Pitched AJVC and Antler, met Next Bharat Ventures, got a referral to F Inc from its founder by cold DM. Stopped in early April.

BDH, a research project. Shut down. An earlier personal site in Astro, with a resume page.

TCCA, the first ARC-AGI-3 agent: 0 to 1 of 7 levels on ls20, then 4 of 183 across all 25 games in two weeks.

TCCA scored 0.263% on the full ARC-AGI-3 benchmark in 147 seconds on a Kaggle CPU. Claude Opus 4.6 and Gemini 3.1 Pro scored 0.2% the same day; Claude's run cost $8,900.

CACCS-AI with Suramyaa Sarkar (University of Oklahoma): narratives into causal loop diagrams, system archetypes, Nash equilibria, simulations. I was the developer.

A digital twin of myself: a private chat with a persona built from my own archive. Fine-tuned a model on my own tweets.

Google × Hugging Face Gemma meetup on 18 April: showed TCCA live to research engineers from Hugging Face and Kaggle.

Founded a research collective with people from Sarvam AI, AI4Bharat, ISI and IIEST. It didn't last.

Applied to F Inc Canopy with TCCA. Talked about a non-invasive BCI on X; a startup wrote in about it.

Interviewed with Entrepreneurs First for its Fall 2026 cohort. I wrote back in June with what I'd built since.

ButterflyGate's first runs on Gemma 4 and Llama. ChromeClaw, ReasonTrace, Casca, Stockfih, within four days. A Windows command center (sbcmd). "to write or not to write?", the first kessays essay.

Sefai. iwm, a symbolic ARC agent; 0 levels, failure analysis written. categorical-cat.

Virgil. ButterflyGate goes public. lifesim, a Monte Carlo life simulator that refuses to predict. ParaReport at the Build with Gemma hackathon (4 Jul). batcave, a private Mac app for goals, fitness and news. DisCo Cohort 001. The register project starts. 32 inbound co-founder requests.

Research collaborator at Jadavpur University. Design partner at Unify; #171 merged. CuteTuner, a local vocal tuning studio. A PR to activity-frames making captured text opt-in, as its README promises. Invited as a design partner by a startup from DeepMind founders. Joined the Thinking About Thinking Ambassador Programme. Selected for PrettiFlow Cohort 2. Researched what it would take to build JARVIS with 2026 technology.

Blindfold at SIH 2026. ResonatorLM audit. Mechanic Composition Worlds, after an ICLR 2027 plan called Solving Is Not Understanding was cut to one paper. maiku. Logbook, a private telemetry app. Blackstar, an ARC Prize 2026 entry. The ML Challenge 2026 entity-resolution entry. The register paper, in prep. Applied to F Inc Blueprint with Virgil; they called the application promising and interviewed me on the 24th. Declined PrettiFlow because of semester exams.

A 70-page monograph on Virgil: "continue the research, narrow the claim". Blindfold's launch video. Started a brain and AI project with a friend; it stays quiet for now. Horizon, a decision workstation for the next 12 months.

On the desk: the register paper, Mechanic Composition Worlds, the brain and AI project, Blackstar, maiku.