02Research & Lab

Research & Lab

Six proposed experiments for after launch. None is built yet, and each card says so.

6 proposed0 built yet

01Candidate projects

Six ideas in orbit.

Each needs a clear user problem, a baseline and an evaluation plan before it is built. Select one to jump to it.

ProposedResearchingBuildingEvaluatedLive
Builds will land on GitHub
AProposed

A small language model on Apple Watch

Research whether a narrowly scoped language capability can run directly on the watch — and say exactly where inference runs if it cannot.

Question
Can a focused language task run on-watch, offline?
Measure
Model size, quantisation, memory, latency, energy use and task quality — compared against a paired-iPhone design.
BProposed

Research-to-experiment agent workbench

Turn a selected paper into a reproducible experiment with explicit tools, citations, run logs and bounded feedback loops.

Question
Can an agent reproduce a paper faithfully, within cost limits?
Measure
Source accuracy, reproducibility, tool success, failures, latency and cost.
CProposed

Graph-grounded knowledge product

A usable assistant over a permitted corpus that compares a retrieval baseline with graph-based retrieval and orchestration.

Question
Does a graph measurably improve answers on a defined task?
Measure
Answer quality, provenance and the effect of graph design — with knowledge-graph and orchestration effects reported separately.
DProposed

Agent harness benchmark

Compare agent configurations on a fixed task suite with trace inspection, bounded retries and controlled changes.

Question
Which configuration changes hold up on held-out tasks?
Measure
Completion, correctness, regressions, latency, cost and reproducibility.
EProposed

A practical local-inference product

Use the Colibri engine in a small local assistant over a public corpus, and benchmark it against a baseline runtime.

Question
How practical are large local models for a real task?
Measure
Model loading, time to first token, throughput, memory, task quality and offline behaviour.
FProposed

Local wearable-data application

A personal wearable-data app with local ingestion, understandable analytics and optional on-device explanations.

Question
Can personal health signals stay on-device and still be useful?
Measure
Observable data flow, raw versus derived measurements, and the usefulness of local explanations.
How research moves
  1. Source
  2. Original evidence
  3. Research note
  4. Hypothesis
  5. Baseline
  6. Evaluation
  7. Demo & write-up

Shipped work is on the Work page.

Production AI and the CaseCraft case study.

See the work