Weco

Building and optimizing machine learning pipelines often involves extensive, manual iteration. Finding the most performant code requires significant time investment in experimentation, tweaking, and evaluation. Weco introduces a systematic approach, leveraging an AI agent to automate this complex process. Powered by AIDE (Agentic Iterative Design Engine), Weco acts as your AI research engineer, turning your evaluation benchmarks into a self-improving system. It autonomously runs experiments, refining your code against your specified metrics to discover higher-performing solutions.

Key Capabilities Driven by AIDE

• 🤖 Automated Experimentation: AIDE systematically generates and tests numerous code variations for tasks like GPU kernel optimization, model development, and prompt engineering, often running hundreds of experiments overnight. This allows you to explore a vast solution space without manual intervention.

• 📊 Metric-Driven Optimization: You define the success criteria – accuracy, speed, cost, or any custom metric. AIDE iteratively refines the code, focusing solely on improving performance against your specific evaluation pipeline until the metric indicates a better solution. This evaluation-driven loop demonstrably outperforms one-shot code generation, as shown in benchmarks like OpenAI's MLE-Bench where AIDE secured 4x more medals than the next best autonomous agent.

• 🧠 Broad Code Search & Domain Understanding: Unlike traditional AutoML limited to specific model classes, AIDE leverages LLMs to search the entire code space. It can apply methods across diverse domains (NLP, CV, tabular) and incorporate background knowledge for specialized fields like finance or biomedical science, often discovering novel approaches humans might miss (validated in METR's RE-Bench).

• 💡 Natural Language Guidance: While AIDE operates autonomously, you can inject your domain expertise or specific requirements using natural language prompts, guiding the agent's search process without needing intricate hyper-parameter tuning.

• 🛡️ Flexible & Secure Execution: Generated code runs within a secure, sandboxed cloud environment managed by Weco, automatically provisioning optimal hardware (CPUs, GPUs). For data sovereignty needs, an on-premises execution option is available, giving you full control. You retain full IP rights to all generated code.

• ⚙️ Language & Framework Agnostic: Optimize code across various languages and frameworks, including Python, PyTorch, TensorFlow, and target diverse hardware from CPUs and GPUs to TPUs and Apple Silicon.

Practical Use Cases

1. Optimizing Existing Model Performance: You have a working PyTorch model for object detection, but latency is too high for production. You provide Weco with your model code, evaluation dataset, and specify the metric (e.g., maximize mAP while keeping latency below 50ms). AIDE iterates on the model architecture, inference code, and potentially quantization techniques, delivering optimized code that meets your performance target.

2. Accelerating Custom Compute Kernels: Your team relies on a custom CUDA or Triton kernel for a critical preprocessing step, but it's becoming a bottleneck. Using Weco, you provide the kernel code and a benchmark measuring its execution speed. AIDE explores alternative implementations, memory access patterns, and parallelization strategies, aiming to significantly reduce runtime, similar to how it outperformed human experts on METR's RE-Bench for Triton kernel optimization.

3. Developing High-Performing Tabular Models: Starting a new project to predict customer churn using a large tabular dataset? Provide AIDE with the data schema description and the target metric (e.g., F1-score). AIDE generates end-to-end pipelines, including feature engineering (potentially crafting domain-specific indicators if guided), model selection, training, and evaluation code, autonomously iterating to find a top-performing solution, often delivering an initial version within minutes.

Why Weco?

Weco offers a distinct approach to ML development by automating the crucial, yet often laborious, optimization cycle. Its strength lies in:

• User-Centric Metrics: Optimization is driven entirely by your data, your evaluation pipeline, and your definition of success.

• Proven Effectiveness: AIDE's capabilities are validated through leading benchmarks (OpenAI MLE-Bench, METR RE-Bench) and adoption by frontier AI labs and research publications (e.g., Sakana AI's AI Scientist-v2). It demonstrably trades compute resources for superior code quality.

• Flexibility: Adaptable to various programming languages, ML frameworks, and hardware targets.

• Scalability: Designed to handle complex optimization tasks over extended periods, continuously refining solutions.

Weco empowers you and your team to focus on higher-level strategy and problem-solving, while AIDE handles the intensive, metric-driven experimentation needed to push the boundaries of your ML pipelines.