Our active research
vectors.
We design, analyze, and build security layers for autonomous agents. Our research targets first-principles vulnerabilities in tool execution, retrieval, and reasoning loops.
Future Directions
Project Hollow Choir
Coordination without instruction.Multi-agent systems research studying how LLM agents develop emergent roles and division of labor without being explicitly assigned them.
Project Amber Room
Memory that doesn't reset.Long-context memory architecture research aimed at making an agent's memory feel continuous rather than retrieved.
Project Fathom
Looking inside small models.Interpretability research focused on fine-tuned small language models, examining what actually changes internally when a base model is adapted via LoRA/QLoRA.
Project Undertow
How models get to the answer.Reasoning trace analysis studying the process models use to arrive at outputs, not just whether the final answer is correct.
Project Halflight
What synthetic data does to a model over time.Studies the compounding effects of synthetic data on model behavior across successive fine-tuning rounds.
Project Marrow
How much intelligence survives compression.Efficiency research measuring how much of a model's capability survives quantization, pruning, and distillation.
Project Nocturne
Intelligence that doesn't need the cloud.On-device and local inference research studying latency, resource tradeoffs, and behavior consistency when running LLMs locally.