Lakuna Router: Making AI Models 'Audition' for Real-World Workflows
Due_Hovercraft6497 · reddit · 2026-07-22
A developer argues that current AI routers are unreliable because they classify prompts as 'simple' or 'complex' before generation, missing niche knowledge or hidden context. To solve this, they are building Lakuna, a router where models must earn their place in a workflow by being compared on actual user tasks.
The project features a detailed roadmap:
- V1 (Personal Comparisons): Connects to local and hosted OpenAI/Anthropic-compatible endpoints. It saves real tasks as a private test set and compares models under identical configurations, using automated tests for objective results and blind A/B testing for subjective quality.
- V2-V4 (Auditions & Routing): Alternative models will quietly 'audition' on real requests without altering the user's received answer. Over time, it maps out which models excel at specific workflows and eventually auto-routes to find an equilibrium between costly frontier models and cheaper alternatives.
- V5 (Model Improvement): Leverages collected data and user preferences to use teacher models to fine-tune weaker models for specific tasks, complete with regression tests.
Currently a side project built by a student, the developer is actively seeking feedback from users with real local-model workflows.
More from coding & agent
- Coding agents could let API providers open migration PRs for breaking changes — ycombinator · 2026-07-23
- Builder shares a 13-file Claude context system that ends repeated explanations — jaindl · 2026-07-23
- Gitlord turns every agent turn into a Git commit and supports MCP — Square_Light1441 · 2026-07-22
- Automatic end-to-end testing now reaches the backend in one integrated flow — blixt · 2026-07-22
- monday.com says AI agents lifted PR throughput by more than 50% — AWS ML Blog · 2026-07-22
- Developers are using Lovable and Google AI Studio for fast UI refreshes — rseroter · 2026-07-22