Challenge
The inputs do not agree with each other. Market data, text, and model output arrive in different shapes, and a signal is only useful if it can survive the move from research into live execution — with latency, failover, and a human still on the risk boundary. The problem was making that path durable, not collecting another dashboard.
Approach
Built ingestion and normalization first, so research and production saw the same time series and text instead of one-off extracts.
Kept a research path separate from live execution, and promoted logic only after it met real operating constraints.
Ran the production side as always-on infrastructure across regions, with human review where risk required it. AI-heavy workflows are used where they compress analyst time.
Outcome
The organization has a system it can keep evolving: messy inputs become a maintained pipeline, research can be promoted into production, and live trading participation is part of how the system is operated. Model and position detail stays unpublished.
Proprietary research and execution infrastructure developed in-house.
Proof
Live systems
Research and production trading paths operated together since 2018. Ingestion, signals, and execution-side integration stay in service.
Details of models and positions are intentionally not published.