I'm Dylan Siegel.
I run QuantDev.ai.
QuantDev.ai is an open-source microstructure alpha lab. The work is to turn raw venue data into predictive market states, build memory over the ones that survive, test whether the signal has information and path economics, and only then package it into a trading system.
Live
Operator
Dylan Siegel
Founder · QuantDev.ai
-
Domain
HFT micro
-
Mode
Open lab
-
Stack
Go · agents
The story
Why I built an open lab.
Dylan Siegel
Founder · QuantDev.ai · Building in public
“Data is not the product. Alpha is the product.”
Most public trading research jumps straight to claims — performance numbers, equity curves, screenshots — without ever showing the sample contract, the label contract, or the evidence behind them. I started QuantDev.ai because that gap is the work. What separates research from storytelling is whether the data was actually asked the question.
The lab is built the other way around. Every study names its sample, its label, and its result. Failed ideas are documented as evidence, not deleted. The output of a research pass is the result — pass or fail.
I focus on raw microstructure because that is where price formation actually happens. Pressure, absorption, replenishment, failed impact, liquidity decay, memory. The goal is not to decorate price with indicators — it is to design predictive market states from the raw behavior of the book.
The public lab is everything you need to do this kind of research yourself: studies, harness code, diagnostics, methodology. The private layer is where surviving research becomes a deployable system — tuned configurations, execution logic, monitoring, and a workspace of serious operators. Cleanly separated.
Focus
What the work is.
Four surfaces — the alpha hypothesis, the public research, the agentic workflow that runs it, and the systems that come out the other side.
Microstructure alpha
I design alphas from raw venue data — market-by-order, depth, trades, quotes. The unit of work is a predictive market state, not an indicator.
Open-source research
Atom studies, memory features, metric stacks, diagnostics, failures. The work is public so the research process is visible — including what did not work.
Agentic engineering
I run the lab with modern coding agents — Codex, Claude Code, Cursor — to move faster through study generation, harness refactors, and failure documentation.
Deployable systems
Surviving research becomes a packaged trading system: signal definition, gates, execution logic, deployment, and monitoring. The algorithm packages the alpha.
Operating principles
How the lab operates.
Six rules the lab follows, in priority order. The rules are what keeps the research honest when an idea looks too good.
- 01 rule
Alpha is the product.
Data, indicators, and infrastructure are inputs. What I publish has to be a claim about market state — or it does not get published.
- 02 rule
Information first. Extractability second. Policy last.
A signal has to predict direction before path. It has to clear cost before it becomes a system. No skipping stages.
- 03 rule
Conclusions over claims.
Every study names its sample, its label, and its result. Failed ideas are documented as evidence, not deleted.
- 04 rule
Raw atom or nothing.
If the raw microstructure atom has no predictive structure, no downstream model rescues it. Memory is built on top of atoms that already work.
- 05 rule
Open in public. Packaged in private.
The research is open. The strongest surviving systems — tuned configs, execution logic, deployment — live in the private layer.
- 06 rule
The research machine is the edge.
One signal is not the moat. The lab — agents, harness, diagnostics, failure log — is what makes the next signal cheaper to find.
Now
Current state of the lab.
-
Working on
Memory-feature studies on BTC perp microstructure
-
Writing about
Path economics — MFE, MAE, first-touch, time-to-pay
-
Stack
Go research harness, agentic study generation, Astro public site
-
Based in
Building in public
Connect
Follow the lab. Run the code.
The research notes are on the site. The harness, study code, and diagnostics are on GitHub. The walkthroughs are on YouTube.