Day 1: Why I'm building formulize.dev
I've been building a GPU-accelerated symbolic regression engine on nights and weekends for six years. Today I bought the domain, shipped a landing page, and started charging money. This blog is the public log of the next two years of that build.
The one-sentence problem
Every scientist, engineer, and quant researcher has messy numerical data they want to turn into a clean mathematical equation. Existing tools — PySR (Julia-backed), Operon (C++), gplearn (2010s-era genetic programming) — are academic-quality, but they're slow, hard to install, and worst of all, they don't tell you how much to trust the equation they return.
I've built something 3-8× faster than PySR with an LLM-typed-program-first proposer and — the part nobody else ships — a bootstrap-structural confidence layer that tells you across 1,000 resamples of your data how often the same equation structure recurs. That last part is the difference between "cool academic demo" and "something a hedge fund quant will bet capital on."
Why now, why me
Two things happened in 2024-2025 that made this the right moment. First, GPUs became cheap enough that a solo builder could run a genuinely competitive engine on a $2K RTX 4090. Second, LLMs (Claude, GPT-5) got good enough to propose typed programs as candidates — which lets us reject 99% of physically-impossible equations in microseconds before spending compute fitting their constants. Miles Cranmer (the PySR creator) said publicly last year that "PySR's LLM proposer is bolted-on; the field needs an LLM-first architecture." I built the LLM-first architecture.
As for why me: I'm not a symbolic-regression academic. I'm a backend engineer who got obsessed with fitting equations to my own experimental data six years ago, tried PySR, hit its limits, and started building. That obsession is the moat — because the field is too niche to attract a well-funded competitor and too specific to be a foundation-model side project.
What "shipped today" means
Right now, "shipped" means:
- The landing page is live with a $99/mo Stripe button.
- The benchmark reproduction script is public — every claim on this site links to a 60-second script that runs it head-to-head against PySR.
- The core engine is open-source at github.com/formulize-dev/formulize under MIT.
- This blog exists, and I've committed to publishing every single day for the next 30 days.
What's not shipped today: a fully-working hosted formulize deploy endpoint. Right now, if you subscribe, I onboard you personally on Slack + WhatsApp and provision your endpoint by hand. That's the Pieter Levels "ship v1 in 24 hours, do things that don't scale" playbook. Automation comes at 50 paying customers, not before.
What I'm not building
Fifteen conversations with VCs, technical founders, and real users over the past month landed on this: the engine is not the problem, focus is. I've spent six years building five different verticals on top of one engine — a fintech signal service, a physics equation tool, a catalysis distiller, a validation tool, a portfolio product. All of them died from lack of shipping, not lack of technology. Today they're all killed. The engine survives, formulize.dev is the only product, and the other verticals become demo templates on the same engine.
I'm also not building: a Rust rewrite (speed isn't the bottleneck), streaming SR (no market), a domain-priors library upfront (wait for user pull), or an enterprise MLOps pipeline (customers will build their own). All of these are seductive because they feel like "real engineering." They're actually procrastination dressed up as work.
The plan for the next 2 years
Ship weekly. Blog daily. Charge from day one. Get to 5 paying customers by end of Week 4 or pivot the pitch. Hit $15-30K MRR by December 2028 while I'm still on an H1B at my day job. Then return to India and decide whether to raise an Indian Series A or keep building indie forever — but that's a decision for month 13, not today.
If you're a quant researcher, materials scientist, biologist, chemist, or physicist who fits equations to noisy data for a living: email me. I'd love to make you our first design partner.
See you tomorrow, Day 2 of 30.