Self-Learning Engineering AI Development
Budget: £250 – £750 GBP
I need a robust, regenerative AI system that can teach itself from the engineering resources I already own. The model must be able to ingest large sets of technical documentation (PDFs, CAD notes, standards, manuals) and simulated test results, then refine its own knowledge base without constant human intervention.
Here’s how I picture the workflow. You design an end-to-end pipeline that reads the documentation and simulation data, fine-tunes on it, evaluates its own performance, and loops the improvements back into the next training cycle. Python with PyTorch or TensorFlow feels natural, but I’m open to other stacks if you can explain the advantage. Reinforcement-learning-from-simulated-feedback, transformer-based architectures, or retrieval-augmented generation are all on the table as long as the final system keeps learning autonomously.
— Required deliverables —
• Source code and training scripts ready to run on my workstation (Linux).
• A demonstration model already trained on a small sample of my documents and simulations.
• Clear documentation that explains the training loop, data preprocessing, and how to add new material.
• A short video or live demo showing the self-learning cycle in action.
The solution must scale; my full dataset is several gigabytes today and will grow. If you can design with modularity so sensor data can be added later, that’s a plus.
When you reply, tell me which techniques you’d apply first and how you’ll prove the model is genuinely improving over each iteration.
Here’s how I picture the workflow. You design an end-to-end pipeline that reads the documentation and simulation data, fine-tunes on it, evaluates its own performance, and loops the improvements back into the next training cycle. Python with PyTorch or TensorFlow feels natural, but I’m open to other stacks if you can explain the advantage. Reinforcement-learning-from-simulated-feedback, transformer-based architectures, or retrieval-augmented generation are all on the table as long as the final system keeps learning autonomously.
— Required deliverables —
• Source code and training scripts ready to run on my workstation (Linux).
• A demonstration model already trained on a small sample of my documents and simulations.
• Clear documentation that explains the training loop, data preprocessing, and how to add new material.
• A short video or live demo showing the self-learning cycle in action.
The solution must scale; my full dataset is several gigabytes today and will grow. If you can design with modularity so sensor data can be added later, that’s a plus.
When you reply, tell me which techniques you’d apply first and how you’ll prove the model is genuinely improving over each iteration.