Lab
AI and system experiments worth publishing.
Where experimental work lands once it is stable enough for a public URL. Three builds are on the bench now. The Lab is not a junk drawer — it is a workshop with labels.
On the bench
Active design and early build. No public demos yet — these entries mark what is coming, not what is finished.
01
Tiny LLM Lab
AI Engineering · Educational · Coming soon
Build the toy model to understand the machine. A first-principles learning lab: train a small, inspectable GPT-style Transformer on a tiny corpus, watch it fail as a useful assistant, then pivot to a transparent retrieval-augmented workflow with a local instruction-tuned model. Glass-box code. Honest limits.
Pending
02
Sentinel Protocol Lab
Agent Security · Systems · In design
The agent proposes. Evaluation decides. A portfolio experiment at the edge where language models meet irreversible external state. The public question is simple: does this exact proposed action fulfill the mandate a human actually signed — or only something that looks close enough?
Pending
03
Project Euclid
Mathematical Discovery · Process · In design
Not a superhuman mathematician — a research process whose value survives every model inside it. Coordinating language models, symbolic tools, computational search, and formal verification so that verified novelty can be claimed only when the record supports it. The unit of progress is the ledger entry.
Pending
What lives here
- Learning labs
- Inspectable builds that teach the machine and the workflow.
- Bounded prototypes
- Narrow demos at real trust boundaries — not general platforms.
- Reference fixtures
- Stable, linkable artifacts other work can point at later.
Most experiments start as code. Follow the repos for updates when Lab artifacts drop.