Model adaptation
Domain adaptation, fine-tuning, preference optimisation, evaluation and distillation.
TECCO LAB / EXPLORATIONSGive the future
Give the future
somewhere to begin.
A place for ideas to become models, experiments and engineering questions.
Browse research programmes01
Experimental findingsSpatial semantics
Room geometry, semantic typing and topology form an editable substrate for downstream device and circuit reasoning.
02
Product architecture proposalBehaviour & environment
Explore the intermediate layer between “I want to read” and the capabilities that can create a suitable environment.
03
Infrastructure directionCloud–edge intelligence
Keep routine control local; route complex reasoning through structured context and return bounded decisions to the edge.
04
Research investigationEnergy observability
Preserve solar, storage and metering channels before attempting explainable attribution and predictive dispatch.
Five workloads.
One governed lifecycle.
Spatial simulation
Home-model development, synthetic environments and policy evaluation.
Agent reasoning
Multimodal context, long-tail requests, memory and planning.
Edge personalisation
Local function calling, quantisation and evaluated household adapters.
Operational validation
Shadow evaluation, rollout checks, resilience and deployment transitions.
Research roadmap · Workloads, model choices and deployment architecture evolve through evaluation.
BUILD WITH USBring a question
Bring a question
worth investigating.
We welcome conversations around spatial models, environment semantics, cloud–edge execution and the evaluation of intelligent systems in real places.
Discuss a lab collaborationStart with our research perspective
Joe Zhou’s essay explains why structured snapshots, context and human corrections form a practical first step towards spatial intelligence.
Read the perspectiveSYSTEMS / IN DETAIL
An experiment needs a falsifiable outcome.
The lab agenda links model adaptation, spatial simulation, agent reasoning, edge personalisation and operational validation. Each workload asks for different evidence; a successful demonstration is a starting point for investigation.
CONCEPTUAL MECHANISM 01 / 04
| Boundary | Responsibility | Engineering consideration |
|---|---|---|
| Spatial modelling | Geometry, semantics and correction cost | Measure the engineering usefulness of the resulting representation. |
| Agent reasoning | Context, permissions and a proposed action | Examine ambiguous requests, conflicting observations and unsupported capabilities. |
| Operational validation | Runtime behaviour and physical feedback | Exercise degraded connectivity and incomplete or stale state. |
Why keep negative results visible?
A failed experiment can change the architecture. The floor-plan investigation found that tiled inference could fragment large rooms; a more impressive local crop was not necessarily a more useful assembled model. Recording the failure and the evaluation conditions helps prevent the same attractive but ineffective approach from returning.