Skip to content

GPU-compiled strategy research

Ask for a strategy.Get ranked evidence in seconds.

TradingFlow compiles your question into a typed dataflow program and sweeps the whole parameter grid on an NVIDIA L4 GPU — charts, rankings, and a reproducible receipt in about a second. Not the minutes it takes a coding agent to write Python.

Illustration of a TradingFlow run: a plain-English question is typed, compiled into a typed dataflow program, swept across a parameter grid on an NVIDIA L4 GPU, and answered with ranked equity curves and a reproducible run receipt. Numbers shown are demo fixtures.

How it works

From question to ranked evidence in three moves.

01

Describe

Your question, in English — tickers, indicators, date ranges, and what “best” means to you.

02

Compile

An agent authors a typed dataflow program. Every node is validated before it touches a GPU — bad programs never run.

03

Sweep & rank

The whole parameter grid executes in one pass on an NVIDIA L4. Top-k comes back with charts and receipts for every answer.

Performance

Measured, not marketed.

Numbers from the production execution path — programs compile once, then the whole parameter grid answers from a warm cache on an NVIDIA L4.

warm query
0.5 s
novel query
1.0 s
code agent
minutes

These bars filled in exactly 0.5 s and 1.0 s — the measured warm and novel query times.

Reproducibility

Every answer shows its work.

res_9f2c41ab77d3

one id pins the data window · the compiled program · the full parameter grid

Each run gets a res_ id that pins the data window, the compiled program, and the full parameter grid. Re-run it next week — same inputs, same answer. Share the id and your team sees exactly what you saw.

Run receipt

run
res_9f2c41ab77d3
program
cross_sweep@41ab
grid
400 combinations
data
NVDA · 71 sessions
hardware
NVIDIA L4
status
▮ reproducible

every run reproducible by id

Start asking better questions.

Sign in and get your first ranked run in minutes — no notebooks, no backtesting boilerplate.