Founder of VARRD. Building tools to find statistically real trading edges with AI.

What VARRD is

VARRD is the governed live edge layer: statistically validated market behaviors, monitored in real time, with sample size, out-of-sample performance, decay, correlation, exposure, and execution context.

"AI has driven the cost of idea generation down to almost zero. Now the bottleneck is different. Now we have to verify them, evaluate them."
— Terence Tao

That's the whole thesis. The trading world has enormous demand for verification, and almost everyone testing ideas with AI is doing it wrong — p-hacking, overfitting, OOS tampering, lookahead bias, no multiple-testing correction. People deploy 25-Sharpe-in-backtest strategies and get wrecked in production. Verification isn't a model problem. It's a context-management, knowledge-graph, and discipline problem.

Two ways to use it:

• Bring your own idea. Traders have insights, but there's a real language and intent barrier between a hypothesis and the quantitative test that proves it. VARRD removes that wall. Describe the idea in plain English; the engine turns it into a properly specified test, runs it under the strictest reasonable statistical settings, and tells you whether the edge is real.

• Skip the research, take the edges. If you'd rather just trade what's already working, the edge feed is there. For a few cents per query, you see which VARRD-validated edges are firing in the market right now, about to trigger on a pending bar close, or already in an open trade. You don't generate the edge — you trade the ones we've verified.

Stuff I've built since I was a kid

• Before 12 — Couldn't play sports for medical reasons, so I learned to fly small prop planes instead. Landed one about 25 times before turning 13. Eventually got a procedure done, went back to sports — didn't love flying enough to stick with it. Just always been a doer.
• 12 — Started an ice cream truck because that's what I'd told everyone I wanted to be when I grew up. First "company," half a joke.
• 13 — Bought a 3D printer. Was one of the first 10 on Thingiverse to grab the fidget spinner model and started printing them along with Pokémon Go gear, sold to kids at my school and a few others. Mom kept the printer fed with filament while I was at school in 6th grade.
• 14 — Saved up for a laser engraver. Cut and engraved wood and granite for paying customers — first real online business.
• 2017 — Stuck Helium (HNT) routers in buildings where I had electricity access, paid owners small amounts for the watt usage. Bought a few Bitcoin miners around the same time. Mostly to learn the rigs, but they made some money too.
• ~2018 — Worked on a venture that captured methane waste from a landfill, converted it to electricity, and used it to power on-site crypto mining. Played a major role in setting it up and running it.
• Bought and flipped domain names (iamlearninghowtogolf among them, minor success). To drive traffic, ran Twitter bots that posted something useful instead of spam — pulled live gas prices via geo scripts and posted nationwide daily averages.
• Got into DeFi early, mostly to learn the mechanics rather than make money on it.
• Building with tensor models since they came out in 2017.
• A dozen other intellectually-curious paths I went down along the way — small ventures, weekend builds — not worth listing individually.

The other side of those years

Worked through my teens as a janitor at a private jet airport, in retail, at a fast food chain, and on a landfill. Hated all of it. Only now do I appreciate why my parents made me do those — they're the reason I'm not letting anything coast.

Career

Traded futures derivatives at one of Chicago's biggest futures firms from 18 to 22. Loved that the job was just you, your data, and your own code — the market eventually told you whether you were right. Still profitable, still trading.

College

Mostly stopped doing the curriculum — the teachers thought I was ahead of what we were covering, so the school let me work on my own ventures instead. They had me run Q&A sessions for the faculty and staff on AI and crypto, and I review their AI and BIS curriculum each year before classes start. Still come back a few times a year to talk to classes.

About this GitHub

VARRD has been built for years on our own servers — most of the work doesn't show up in this commit graph. What's here is the public surface (CLI, MCP, SDK).

VARRD — how we got here

Started this with Ben (Princeton math, buddy from Chicago). My father — one of the most successful derivatives traders in North American history — works hand in hand with us on the big-picture direction. Four phases:

1. Tools for ourselves. Built research tools for ourselves and the Chicago futures firm we were trading at. Just trying to find edges faster.

2. The engine. People were pointing AI at price data, getting back strategies with 25 Sharpe ratios, and getting wrecked when they deployed them. The AI wasn't lying — it was falling into all the invisible traps that kill real quant research: p-hacking across thousands of variations, OOS tampering, no K-tracking (every test should raise the significance bar — AI tools didn't track that), lookahead bias, ungated data normalization, no multiple-testing correction. We built a system that takes any plain-English idea, turns it into a properly specified quantitative formula, and tests it under the strictest reasonable statistical settings — every guardrail enforced at the architecture level so even an AI agent can't skip them. Took a couple years to build the DSL and the surrounding system. The rule baked in: transparency, legitimacy, and honesty about what's actually being tested comes before any individual result. Most ideas get killed. The ones that survive are real.

3. Autonomous mode. Realized that fully driving the engine took more bandwidth than most traders had time for. Built a mode where you point VARRD in a direction and it goes off and tests, discovers, and surfaces edges on its own — same context management, same test rules.

4. Edges as a service. People don't want a drill, they want a hole in the wall. Years of testing — by us, our pilot users, and now hundreds of researchers — built up a real library of edges, and we have meaningful compute through NVIDIA Inception. Privacy was a transparent paid option, taken by users who wanted it; everything else aggregates into a live feed. Now you can pay a few cents per query to see which VARRD-validated edges are firing right now, about to trigger on a pending bar close, or already in an open trade. Across the last 70 days the feed's edges have made money with small drawdowns. Edge discovery stopped being the bottleneck a while ago — properly attained edges, portfolio execution, and risk management are. So we cut out the first one.

With the agent economy ramping up, the natural interface is your AI talking to ours — paying small fees to query the research engine, autonomous mode, or the live edge feed. That's where we've leaned in.

Hundreds of users today, thousands soon at our pace. 100,000+ hypotheses tested. MCP server shipping now: pip install varrd.

It's been a ride and we're nowhere near done. What VARRD is today is a fraction of what it'll be. The early signs of the impact we set out for are already showing up.

Where this is now

• NVIDIA Inception partner
• Built alongside Wedbush's business development team
• Direct conversations with C-suite leadership across 30+ major financial firms — clearing brokers, prop trading shops, brokerage tech
• Pilot programs with 20+ brokerages, including most of the major US clearing firms
• Used by major hedge funds and prop trading firms
• Context-engineering work has caught unsolicited interest from major Silicon Valley AI labs — stayed focused on VARRD
• Multiple acquisition discussions active — holding out for terms that match where we believe this is going
• Bootstrapped — multiple funding offers fielded, the system funds itself, no equity given up

Reach

@VarrdInc · varrd.com