Research

Research that makes AI work in production

We research transformer models, inference, and agent reliability, and prove ideas with PoCs before they reach production.

Public Lab Notes

We separate promising ideas from systems that actually ship

Some research and PoC work stays private under confidentiality.

Pre-training

Focus: Transformer architecture, data quality, and training efficiency.
Method: We design the architecture and data, then train at scale with 5D parallelism.
Signal: Models built for the task, not borrowed and forced to fit.

Fine-tuning

Focus: Domain behavior, tool use, and alignment.
Method: We fine-tune any model alongside regression checks and task scorecards.
Signal: Changes you can measure before release.

Inference

Agents & RAG

Focus: Tool use, retrieval quality, and loop control.
Method: We replay traces, test permissions, and tune harness and loop behavior.
Signal: Agents that stay useful and in control.

Model & inference research

Check 01

How transformer architecture choices affect quality and cost.

Check 02

How 5D parallelism makes large-model training and serving efficient.

Check 03

How serving limits should shape model choice before the build.

Reliability

Check 01

Evaluation across quality, latency, safety, and control.

Check 02

Reliability testing for tools, planning, and escalation.

Check 03

Release criteria tied to governance and deployment needs.