R-01
Retrieval that holds at corpus scale
Most retrieval quality loss happens at ingestion, not inference. We study structure-aware chunking, hybrid lexical-semantic scoring and freshness decay on corpora above a million documents.
Active
Research lab
A lab with a delivery obligation.
We do not publish papers for their own sake. Every thread here exists because a client system was not good enough, and every finding ends up in a product within a quarter or two.
How the lab works
Three rules that keep research honest.
Problems come from production
A thread opens when a real deployment hits a ceiling — retrieval quality, latency, thermal envelope, reliability under bad tools.
Findings must be reproducible by the delivery team
If an engineer on a client engagement cannot apply it within a sprint, it is not finished research.
Negative results are published internally
The approaches that did not work are the most valuable asset a small lab has. We keep them, and we tell clients about them.
Active threads
What we are working on right now.
R-02
Distillation for constrained hardware
Compressing task-specific capability into models that fit the power and thermal envelope of a board we can actually manufacture, without falling off an accuracy cliff.
Active
R-03
Document structure understanding
Reading order, nested tables, stamps, handwriting and multi-language European forms — treating a page as a spatial structure rather than a stream of tokens.
Active
R-04
Agent reliability under real tools
How multi-step agents fail when tools are slow, partially available or wrong — and which supervision patterns recover without inventing an answer.
Active
R-05
Evaluation as a first-class artifact
Treating the eval set as the deliverable clients keep. Cheap, honest, adversarial suites that catch regressions before customers do.
Ongoing
Research summaries are illustrative descriptions of internal work; experiment notes are shared under NDA.
Hardware co-design
The lab has a machine shop behind it.
Distillation targets are chosen against a board Xelec can actually manufacture, with a measured power and thermal envelope. Most labs guess at deployment; we validate it on the bench.
5
Active research threads
< 1 qtr
Typical finding to product
60k
Frames in a distillation set
2
Reference boards in validation
Next step
Have a problem that needs research, not integration?
If your use case sits beyond what off-the-shelf models do, that is the work the lab exists for.