[01] ops
Autonomous agents
Agent systems that plan multi-step workflows, evaluate their own outputs, and operate under human oversight. Humans drive prioritisation and interpretation. Mechanical work is delegated.
◢
Entry 001 — The lab is the experiment
Relaxed Constraints.
A lab wired to improve itself: AI runs as much of the research loop as possible, and every trace feeds the next iteration.
Current Focus
// pillars.md[02] loop
Self-improving systems
Every agent action and routing decision produces a trace. Those traces are data: analysed by other agents, fed back into process. The lab learns from its own behaviour.
[03] ML
ML & data systems
Neural architectures for semi-structured data — density estimation, representation learning, and generative models that work natively with nested JSON. Applied to cardinality estimation, query optimisation, and learned indexes.
Research Blog
> allApr 28, 2026
Breaking Through Tabular Constraints for Synthetic Data Generation
Most synthetic data generation tools assume flat tables. Real-world application data is often nested JSON with optional fields and variable-length arrays. The ORiGAMi architecture handles semi-structured data directly.
Mar 19, 2026
Loosening the Reins
AI is moving so fast that we're becoming the bottleneck. On LLMs as idea search engines, the reversal of control, and where this all goes.