Want a reinforcement learning agent that actually runs in production - not just a notebook demo? I design custom Gymnasium environments, train PPO/TCN agents in PyTorch, and compile them to ONNX for sub-second inference on cheap CPU servers.
WHAT YOU GET
- Custom Gymnasium 1.4 environment: state/action spaces, reward shaping, seeding and reproducible resets.
- Trained PPO agent (or DQN/A2C to fit your dynamics) with checkpoints and full training curves.
- ONNX Runtime compilation for millisecond CPU inference - no expensive GPU servers required.
- Backtesting / benchmark report comparing against a baseline strategy.
- FastAPI harness and Dockerized, reproducible setup.
- 50+ unit tests on the production tier.
PACKAGES
Basic - Environment Prototype: custom Gymnasium env, state-action definition, reward stub. Delivery 3 days.
Standard - Trained Agent & Eval: full environment + PPO/TCN training + backtesting report. Delivery 7 days.
Premium - Production Quant Engine: environment + trained agent + ONNX compilation + FastAPI + 50+ unit tests. Delivery 12 days.
FAQ
Q: Do I need expensive GPU servers? A: No. I compile trained PyTorch models into optimized ONNX graphs for sub-second inference on standard CPUs.
Q: Which RL algorithms? A: Primarily PPO and TCN for time-series/simulation, plus DQN/A2C depending on your problem.
Q: Can you prove RL experience? A: Yes - my open-source Intelligences-Trader has 151 passing tests and production ONNX integration.
Send your problem definition, available data and success metric, and I will propose an environment design within 48 hours.
Proof of work: Intelligences-Trader - github.com/amir-hdri/Intelligences-Trader (151 tests, Gymnasium 1.4, ONNX Runtime).