Research

The AI ecosystem is evolving away from a monoculture of giant frontier models toward speciation. While frontier labs build massive "Oracle" models trained on everything, we're witnessing an explosion of specialized, smaller models — optimized for math, biology, edge devices, or coding.

Just as the animal kingdom evolved different brains for different ecological niches, AI will have purpose-built models for specific domains. A 7B parameter model fine-tuned on radiology outperforms a 70B generalist on chest X-rays. A code model trained with execution feedback writes better functions than one trained on next-token prediction alone.

We're building the infrastructure to train and deploy these specialized species — the tooling, the evaluation harnesses, and the orchestration layers that let them work together as a coordinated ecosystem.

True auto-research — where AI models are the primary drivers of creating better AI models. This isn't science fiction; it's the next phase of ML engineering. We're building Director-agent systems where one agent reads papers, another generates implementation plans, another writes code, and another reviews it.

These loops run 24/7, systematically exploring the search space of model architectures, training recipes, and data strategies. The human role shifts from doing the research to setting the reward metric and reviewing breakthroughs.

The compounding effect is real: each cycle produces marginally better models that are marginally better at running the next cycle. We're investing heavily in the safety and evaluation infrastructure to keep these loops legible and controllable.

Current LLMs forget everything between conversations. Every interaction starts from zero. We're researching persistent memory architectures — context graphs that give agents true long-term memory, relationship awareness, and the ability to build on previous work.

Think of it as moving from stateless inference to stateful intelligence. An agent that remembers your codebase architecture, your team's preferences, your past decisions — and uses that accumulated context to make increasingly better suggestions over time.

Our approach combines episodic memory (what happened), semantic memory (what things mean), and procedural memory (how to do things) into a unified graph structure that agents can query, update, and reason over.

Most voice AI is a pipeline: STT then LLM then TTS. Three separate systems duct-taped together, each introducing latency and losing signal. We're researching native voice-reasoning models where understanding and generation happen in a unified architecture.

Models that don't just hear words — they understand tone, hesitation, emphasis — and respond in kind. A customer saying "that's... fine" with a falling tone means something very different from "that's fine!" with rising energy. Current pipelines lose this entirely.

Our work focuses on end-to-end architectures that process audio tokens directly, reason over them, and generate spoken responses — all within a single forward pass. The result is lower latency, richer understanding, and more natural conversation.