← All projects

Machine Learning Platform

CERES

A research-to-production trading platform that unifies market-data ingestion, model training, walk-forward backtesting, live execution, risk controls, and real-time monitoring.

PythonFastAPIPostgreSQLReactDockerWebSocketsXGBoostLightGBM
Active
~16Kapplication LOC
16backend interfaces
396model artifacts
11configured assets

01 · Problem

What the system solves

Quantitative research systems often fail when backtest logic and live execution diverge. CERES centralizes decision logic so identical feature, regime, and risk rules drive both environments.

02 · Architecture

How the pieces connect

01Exchange APIs
02Feature pipeline
03Regime models
04Decision engine
05Execution + risk
06Live dashboard

03 · Technical highlights

Implementation details

01

Shared decision engine powers both live execution and canonical walk-forward backtesting.

02

Bull, bear, and range-specific XGBoost and LightGBM ensembles route decisions by detected market regime.

03

FastAPI exposes 15 REST endpoints and a WebSocket stream for positions, signals, trades, and market state.

04

A supervised local stack launches 10 trader processes with locking, watchdog recovery, and Dockerized PostgreSQL.

04 · What this demonstrates

Engineering signal

Backend system design across APIs, persistence, background processes, and operational tooling.

Applied machine learning with repeatable training, calibration, inference, and evaluation workflows.

Production-minded safeguards including dry-run operation, circuit breakers, reconciliation, and failure recovery.

Continue exploring

View all projects