Lakuna
Lakuna
Chinmayi Ramasubramanian is a Princeton computer science student with research experience in reward signals, reinforcement learning fine-tuning, and understanding model behavior.
Chinmayi Ramasubramanian is a Princeton computer science student with research experience in reward signals, reinforcement learning fine-tuning, and understanding model behavior.
Meet
Meet
Chinmayi
Chinmayi
Building the personalized audition and routing layer for local, open, and frontier AI.
Building the personalized audition and routing layer for local, open, and frontier AI.
Frontier models can now solve Olympiad-level mathematics and contribute to research problems. That is extraordinary, but most everyday work does not require frontier intelligence. Smaller proprietary models, open-weight models, and models running locally on ordinary hardware can already handle much of that work remarkably well. Yet most people still default to Claude or GPT.
In user conversations, the reason was simple: trust and friction. Choosing models manually interrupts the work and forces people to move context between systems. Automatic routing removes that friction but introduces an invisible risk. Engineers at large technology companies said that even when internal routers reported excellent overall performance, they were barely used because one unexpectedly weak answer could derail an entire workflow. Preliminary simulations also showed that, across four common user profiles, the model selected for an individual matched the global router’s choice only 27% of the time. The theoretically best model for the average user is often the wrong model for a particular person.
Lakuna approaches routing as an audition. Candidate models perform the same kinds of tasks a user already gives their most trusted model. Checkable outputs are evaluated objectively, while blind comparisons capture qualities that require human judgment. When a smaller or local model proves that it can match or outperform the trusted frontier model for a particular cluster of work, it earns that traffic. Lakuna continues checking its performance and routes back to the frontier when a request falls outside what the candidate has proved it can handle.
The future of AI will be hybrid: frontier models for work that genuinely requires them, with smaller, specialized, open, and local models handling everything they can do just as well. Lakuna learns from what each user accepts, rejects, revises, and prefers, applying the logic of recommendation systems to model routing. Over time, the same evidence can become training signal that helps local and open models serve that user better.
Lakuna is opening its first alpha to everyday users and is in early pilot conversations with teams in trading and fintech that already use multiple models but cannot tolerate an invisible loss in quality.
Frontier models can now solve Olympiad-level mathematics and contribute to research problems. That is extraordinary, but most everyday work does not require frontier intelligence. Smaller proprietary models, open-weight models, and models running locally on ordinary hardware can already handle much of that work remarkably well. Yet most people still default to Claude or GPT.
In user conversations, the reason was simple: trust and friction. Choosing models manually interrupts the work and forces people to move context between systems. Automatic routing removes that friction but introduces an invisible risk. Engineers at large technology companies said that even when internal routers reported excellent overall performance, they were barely used because one unexpectedly weak answer could derail an entire workflow. Preliminary simulations also showed that, across four common user profiles, the model selected for an individual matched the global router’s choice only 27% of the time. The theoretically best model for the average user is often the wrong model for a particular person.
Lakuna approaches routing as an audition. Candidate models perform the same kinds of tasks a user already gives their most trusted model. Checkable outputs are evaluated objectively, while blind comparisons capture qualities that require human judgment. When a smaller or local model proves that it can match or outperform the trusted frontier model for a particular cluster of work, it earns that traffic. Lakuna continues checking its performance and routes back to the frontier when a request falls outside what the candidate has proved it can handle.
The future of AI will be hybrid: frontier models for work that genuinely requires them, with smaller, specialized, open, and local models handling everything they can do just as well. Lakuna learns from what each user accepts, rejects, revises, and prefers, applying the logic of recommendation systems to model routing. Over time, the same evidence can become training signal that helps local and open models serve that user better.
Lakuna is opening its first alpha to everyday users and is in early pilot conversations with teams in trading and fintech that already use multiple models but cannot tolerate an invisible loss in quality.