San Francisco
I’m Arvin Bhangu. I build Roam: spatial AI infrastructure that turns the physical world into a persistent, queryable record, so agents and robots can reason about places the way they already reason about text.
Before Roam I built the first uncensored large language model, trained from scratch, tokenizer through serving stack. It reached 1.2 million people in 30 hours. Before that, a mortgage company underwriting against bitcoin, and a decade of shipping software and hardware starting at 14.
Outside the companies
Superintelligence, February to December 2023. Raw text in, public product out.
Read the buildTrained the tokenizer rather than borrowing one, so vocabulary matched the corpus.
Ingestion, cleaning, deduplication, filtering, versioning. Most of the work sits here.
Decoder-only transformer: attention blocks, positional embeddings, normalisation, feed-forward.
PyTorch on NVIDIA clusters. Mixed precision, gradient accumulation, checkpoint recovery.
Instruction following, without collapsing back into the refusal behaviour it was built to avoid.
Perplexity, reasoning, coding, hallucination rate, adversarial prompts, regression between checkpoints.
KV caching, batching, quantization, streaming generation, sampling controls behind an API.
Autoscaling GPU capacity and holding uptime while 1.2 million people arrived in 30 hours.
Why I build
Almost everything I want for people is bottlenecked by thinking. The cure that sits one experiment away from being found. The question a kid asks once, gets nothing back for, and never asks again. All of it is a shortage of attention, and attention is exactly what intelligence makes cheap.
Right now good thinking is expensive and it goes to whoever can pay. I want it to cost close to nothing, the way electricity did once the grid was built, so the person with the hardest problem and the smallest budget gets the same help as anyone else.
I want people to be prosperous. That is what general intelligence is for, and it is why I give my life to it.
Each one set up the next.
The tokenizer, the pretraining corpus, the transformer, the training runs, the evals, and the serving stack were all built here.
Whoever trained the model decides what it refuses. Everyone downstream inherits the decision.
Property, bitcoin, and what happens when software has to live with regulation instead of talking around it.
If bitcoin is real money, it should be able to move through real property workflows without forcing liquidation or trust theater.
I want my work to add up to artificial general intelligence. Roam exists because an intelligence that cannot perceive, remember or reason about physical space is not general.
Software I shipped between 2009 and 2021 reached more than 25 million people, before any of the current work.
Most of what I believe about building for the physical world came from watching infrastructure work, and break, in places with very different constraints.
A private licence. The checklist habit came home with me to engineering.
Built for fun
Optics and capture, the input side of every spatial system.
Application-specific silicon and the thermal and power design around it.
Power delivery for devices expected to run unattended.
Generation and storage, sized against real load.
Control loops, telemetry and radio links.
Local storage and GPU boxes built to train and serve models without renting them.