Sep 2025 — Aug 2026 · New York
Stealth fintech
- <100ms
- from the Kraken feed to the order book
- 90%
- less morning analysis with the pre-market brief
Kraken futures ingestion
Python and Kafka on the Kraken feed, under 100ms.
- Problem
- They needed a live feed going into the trading book.
- Constraint
- Kraken crypto futures, and under 100ms.
- Decision
- Python and Kafka on the official feed, with agents on top of the book.
- Outcome
- Ticks come in and orders go out.
Forecasting and retraining
PyTorch models that retrain around the clock.
- Problem
- They wanted signals that kept up with the market.
- Constraint
- Training had to continue throughout the day.
- Decision
- PyTorch time-series models that retrain all day.
- Outcome
- Live signals and backtests, with models that retrain on their own.
News processors
News text in, pivot flags out.
- Problem
- Some moves show up in the news before they show up in prices.
- Constraint
- The input is live headlines.
- Decision
- A news processor running on the same schedule as the forecasting job.
- Outcome
- Pivot flags from headlines, alongside the time-series signals.
Portfolio audit and pre-market brief
Agents on the SQL ledger, and a morning brief that cut the desk’s analysis work by about 90%.
- Problem
- The morning portfolio check was all manual SQL.
- Constraint
- The models had to work with the ledger they already used.
- Decision
- Agents that query the SQL, plus a PyTorch-based brief before the open.
- Outcome
- They run the brief every morning.