firelex/jeff
Millisecond decisions, any domain: a 0.8B open "System 1" model that picks between your options with calibrated probabilities. One base, swappable LoRA adapters, on your own hardware.
What it solves
Jeff simplifies the process of making automated decisions based on unstructured data. Instead of writing complex logic to interpret text or inputs, you can treat specific decision-making tasks as simple, typed function calls that return a clear result.
How it works
Jeff uses a library of specialized "decision functions"—pre-trained adapters that answer narrow questions. You pass unstructured data into these functions, and they return a result with a "calibrated confidence" score. If the system isn't confident enough in its assessment, it can abstain from making a decision. You can run these functions locally on your own hardware (using Python, Node, Rust, C#, or Java) or via a hosted API.
Who it’s for
Developers who need to integrate reliable, automated decision-making into their applications—such as safety checks, content moderation, or input validation—without building and maintaining custom machine learning models from scratch.
Highlights
- Typed Results: Returns structured, predictable output rather than raw text.
- Calibrated Confidence: Provides a confidence score for every decision, allowing the system to abstain when uncertain.
- Flexible Deployment: Supports both local execution (on Apple silicon or NVIDIA GPUs) and a managed API.
- Multi-Language Support: Includes SDKs for Python, Node, Rust, C#, and Java.
- Narrow Focus: Each function is designed to answer one specific question, making it easy to integrate into existing workflows.
Related
- Dispatch
- Dispatch
- Dispatch
- Dispatch
- Dispatch