Home/Services/Micro AI Research
Service

Micro AI Research and Model Training

Small models, expertly tuned — specialists that beat generalists on the tasks that matter.

Micro AI Research and Model Training
<10B
Parameters, matching yesterday's frontier
5 modalities
Vision · RF · acoustic · spatial · language
10M → 30B
The model-size ladder, tiered to hardware
Single-digit ms
Distilled students on the hot path
Overview

The opposite instinct

The industry's instinct is to reach for the largest model available. Ours is the opposite. By 2026, a well-tuned sub-10-billion-parameter model routinely matches yesterday's frontier models on the narrow task it was trained for — at a fraction of the compute, on hardware you can hold in one hand, inside a power budget a battery can sustain.

Our bet, proven across our deployed platforms: specialisation plus composition beats generalisation — many small experts, each owning one modality or one task, coordinated by a larger reasoner, with the final verdict always committed by deterministic code, never by a model.

    What we deliver
  • Sensor- & domain-specialist micro-models
  • A disciplined tune-quantise-distill pipeline
  • Right-sizing across the model-size ladder
  • Anti-poisoning gates on every training corpus
  • A governed, signed, drift-monitored model lifecycle
The specialists

One specialist per modality

Each tuned to its telemetry vocabulary and served where the data is.

Vision

Tiny object detectors for the hot perception path, and compact vision-language models for scene reasoning, captioning, and visual diagnostic flags.

RF & signals

Task-specific classifiers on raw IQ streams and spectrograms for modulation recognition and emitter characterisation.

Acoustic & seismic

Lightweight signature classifiers for waveforms and spectrograms, from urban soundscapes to unattended ground sensors.

Spatial

LiDAR and point-cloud perception for mapping, tracking, and change detection.

Language & documents

Small language models for report understanding, structured extraction, and tool-calling agents.

The method

We don't pre-train foundation models. We make strong open bases excellent at your task.

01 · Base

Start strong

Begin from the strongest open-weight base for the job — never from scratch.

02 · Adapt

Tune efficiently

Parameter-efficient fine-tuning (LoRA / QLoRA) keeps the adaptation small, auditable, and cheap to re-issue as the mission changes.

03 · Quantise

Fit the tier

Quantised to the deployment tier (AWQ, GGUF, INT4) to fit its memory and power envelope.

04 · Distill

Shrink the hot path

A tiny student distilled from a larger teacher where the hot path demands single-digit-millisecond inference.

The model-size ladder — right-sizing is the research problem
10M – 500MTiny models on the smallest nodes
100M – 10BMicro & small language models on standard and heavy edge hardware
10B – 30BMedium models at cluster hubs
30B +Larger reasoners only where the fabric can carry them
Trustworthy by training

A model is only as defensible
as the data that shaped it.

Every training corpus passes an anti-poisoning ingestion gate: provenance-checked sources, integrity controls, and documented lineage. Performance is gated against curated, gold-standard evaluation datasets — measured on disaggregated subgroups rather than flattering averages.

In our framework, a model is a form of policy encoding. Every model we ship is versioned, signed, and entered into a chain of evidence; every deployment is reproducible; and drift monitoring runs alongside operational metrics so silent degradation is caught, never discovered. The same lifecycle discipline that governs the rest of the platform under Tangram and Origi.

Where we apply it
Defence intelligence disciplinesImagery · signals · acoustic Urban sensing modalitiesVideo · RF spectrum · environmental Medical signal analysisVitals · respiratory · visual flags Legal & policy documentsUnderstanding · extraction
Enkidu

Bring us one modality

And a few thousand labelled examples — we'll show you what a specialist can do. Every technique we recommend is one we have fielded on deployed edge hardware.

Start a specialistTalk to the team