Launched this week
FINAI by Synlabs
Ultra-low latency predictive ML layer
4 followers
Ultra-low latency predictive ML layer
4 followers
FINAI 2.0 — an ultra-low-latency predictive ML layer for markets. Feature engineering → training → strict no-look-ahead sim → inference, in Rust/C++ at 4–50 µs/tick, enabling tick-level backtesting. Per-bar signals (probabilities + raw) across tickers & timeframes, asset-agnostic, strictly causal. Transparent by design: live public board + immutable S3 audit trail — every signal timestamped, independently verifiable. For market makers, execution desks & systematic funds.
Hey Product Hunt! 👋 Nikita here, CEO of Synlabs.
We built FINAI 2.0 out of frustration: serious research on market microstructure needs tick-level simulation, and Python just can't run it at scale. So we rebuilt the whole pipeline — feature engineering → training → strict no-look-ahead simulation → production inference — in Rust/C++, at 4–50 µs per tick. That speed is the unlock: we can backtest at tick resolution, honestly, with no look-ahead.
The engine emits per-bar predictive signals across assets and timeframes — from crypto ticks to US equities. The part we care about most: you don't have to trust us. There's a live public board with a real-time signal stream and an immutable S3 audit trail — every signal timestamped and independently verifiable.
We're a small team a bit obsessed with causal, no-leakage ML — we publish only what survives out-of-sample. Would genuinely love feedback, especially from anyone who's fought the same latency / look-ahead battles. Happy to answer anything 🙏
— Nikita
Plugged FINAI into some of my own tick data and the latency numbers actually held up, which honestly surprised me. The public audit board is a nice touch too, kind of rare to see that level of transparency in this space.
@feride188591 Hi Feride! Thank you so much for testing and for the great feedback.
I'm very glad to hear that the latency numbers match your own tick data! We've invested significant engineering effort into optimizing the native Rust execution path to eliminate any garbage collection or serialization overhead.
Even though the board on the demo server is currently slower than the testbeds, the latency still allows us to verify causality.
Regarding a public audit board, we felt it was absolutely necessary. The quantitative segment is plagued by over-reported backtests and forecasting biases, so we wanted to prove our out-of-sample metrics in real time.