Real-Time FPGA-Accelerated Crypto Trading System

I led a public multi-contributor team repository building a low-latency system that consumes live Binance order book data, computes features from it in hardware, and predicts short-horizon direction.

Feature pipeline

I engineered the real-time feature pipeline in the team codebase, computing 17 order book microstructure signals over a two-second horizon. At this horizon the useful information is in the shape of the book rather than the price series: imbalance, depth and how quickly each side is being consumed.

Online learning

I contributed an online-learning module using incremental outer-product updates, which allows the model to be updated live, in production, without retraining from scratch or taking the system offline. This matters when the data-generating process shifts faster than a retraining cycle.

Hardware acceleration

I benchmarked the FPGA acceleration path at roughly 18x average speedup versus NumPy, peaking at 68x, with 57.6% live directional accuracy. The gap between average and peak is itself informative: it shows where the pipeline is bound by data arrival rather than by compute.