CTF Capital

Where they were

CTF Capital is a crypto investment fund that runs algorithmic trading strategies. They had a strong quantitative research team: people who understood financial models, signal generation, the Triple Barrier Method. But everything lived in notebooks. There was nothing that could ingest real-time Binance data, compute features, generate predictions, and feed a trading bot at production scale. The gap between “the model works in my notebook” and “the model trades real money” was massive.

What we found

The problem wasn’t a model. It was the complete absence of an ML lifecycle. No experiment tracking, no model versioning, no drift detection. A proprietary dependency (mlfinlab) was creating vendor lock-in on core algorithms. And the system needed to operate in two simultaneous modes: batch (daily OHLC aggregations from Binance, stored in Delta Lake) and real-time (streaming predictions feeding a live trading bot).

What we built

A real-time prediction pipeline: OHLC → Features → Filtering → Prediction → Bet Sizing → Strategy. Over 20 financial features: RSI across 5 periods, Ultimate Oscillator in 3 configurations, Chande Momentum Oscillator, Parkinson and Corwin-Shultz volatility estimators, EWMA moving averages. Full ML pipeline on SageMaker (preprocessing, optimization, evaluation, training, deployment, and monitoring), all automated. MLflow for experiment tracking. We migrated mlfinlab to a custom library (ctf-lab) to eliminate vendor lock-in. 12+ AWS services orchestrated: S3, SageMaker, EC2, ECS, ECR, SNS, SQS, Lambda, RDS, DynamoDB, CloudWatch, Airflow. Monitoring stack on Grafana: quality drift, feature drift, importance drift, Slack alerts, and automatic retraining triggers.

What they can do now

8x faster model development and deployment. Production-grade backtesting with real metrics: Sharpe Ratio, Sortino Ratio, MaxDrawDown, Total Profit. The platform runs autonomously: monitoring catches degradation, triggers retraining, and alerts via Slack. Zero vendor lock-in: CTF owns every line of code, including core algorithms. The infrastructure we built was part of what made CTF acquirable by Borderless.

Why they stayed

28 months. deployr’s longest engagement. It started as pure ML engineering and expanded into monitoring, observability, and new data source integrations. There was no single long contract. Multiple scopes stacked up because we earned trust with working code, not slide decks.