Lakuna
Lakuna
Chinmayi Ramasubramaian comes from ML research in reward signal design, RL fine-tuning, and model interpretability.
Chinmayi Ramasubramaian comes from ML research in reward signal design, RL fine-tuning, and model interpretability.
Chinmayi Ramasubramaian
Chinmayi Ramasubramaian
Princeton
Princeton
Managed continual learning for open-weight LLMs: turn your production data into a model that gets better every week, on infrastructure you control.
Managed continual learning for open-weight LLMs: turn your production data into a model that gets better every week, on infrastructure you control.
74% of enterprise AI agent deployments are being rolled back, while 98% of enterprises report increasing AI investment. The companies hitting this hardest can't use commercial APIs: regulation forbids their data leaving controlled boundaries, or the data itself is the IP. They run open-weight models with rich production signal and no way to use it. Prompts plateau; one-time fine-tunes go stale.
Lakuna ingests the signal a company already produces, PR reviews and merge outcomes for coding agents, conversation logs and supervisor labels for customer service agents, and continuously fine-tunes an open-weight model on the company's own infrastructure. A hard constraint layer enforces business rules at inference time regardless of what the model learned, so improvement can be aggressive without risking a catastrophic output.
A side-by-side demo against base-model and retrieval baselines is in progress; regulated-industry design partners are the next step.
74% of enterprise AI agent deployments are being rolled back, while 98% of enterprises report increasing AI investment. The companies hitting this hardest can't use commercial APIs: regulation forbids their data leaving controlled boundaries, or the data itself is the IP. They run open-weight models with rich production signal and no way to use it. Prompts plateau; one-time fine-tunes go stale.
Lakuna ingests the signal a company already produces, PR reviews and merge outcomes for coding agents, conversation logs and supervisor labels for customer service agents, and continuously fine-tunes an open-weight model on the company's own infrastructure. A hard constraint layer enforces business rules at inference time regardless of what the model learned, so improvement can be aggressive without risking a catastrophic output.
A side-by-side demo against base-model and retrieval baselines is in progress; regulated-industry design partners are the next step.