Exigence — applied intelligence

Four places the work sits.

A camera and lidar mast on open moorland under low cloud. — Moorland range. Camera and lidar, no crew in frame.

Defence sensing

Defence-science programmes. On-edge perception with thermal, lidar, and hyperspectral sensors. Cross-domain matching between surveillance imagery and document photographs. Stand-off optical measurement. Pan-tilt and lidar fused into world coordinates. Prototypes taken onto FPGA and ASIC-class hardware, and a portable RGB plus long-wave infrared capture rig built to a tight specification. UK security clearance is available for this work. Written experimental results went to the sponsor.

An evening records room. Monitors show abstract segmentation of equipment, rain on the window. — Records room. Segmentation studies, documents turned away.

Public institutions

Technical diligence for a major administrative and financial institution, invited in by a top-tier firm. Privacy-preserving vision for local-government partners: demographic signal retained, identity removed. Clinical sensing in the home, so a medical reading stays attached to the right resident, including through a change of clothes. The product is a judgement and a model, not a slide over someone else’s notebook.

Dawn aerial of hedged farmland, one field rendered in magenta false colour. — False-colour pass across mixed farmland at dawn.

Land and agritech

The instrument stack from the defence and care work, pointed at terrain and crop. Hyperspectral and thermal passes for canopy and soil contrast. Edge motion models where a camera has to run without a rack. Simulation when a field trial is the expensive part. Engagements are taken on methods already delivered, applied here to land.

A small passenger jet in a dark wind tunnel, pale smoke streaming over the nose and windows. — Smoke visualisation on a regional airframe.

Simulation and production AI

Learned turbulence models for commuter-aircraft shape, trained on ground vehicles and transferred blind onto subsonic hulls. Lift, drag, fuel, and aeroacoustic noise held as objectives. Elsewhere: panoptic segmentation, vision transformers, multi-object tracking, and agentic tooling inside a visual-effects pipeline, so occlusion stops breaking a rotoscope sequence. Current work is language-model systems: agents, memory, retrieval, evaluation, and workflows that survive contact with a real corpus.

Anonymised

Selected engagements.

01 Diligence

Before the data room

Invited, via a top-tier consultancy, to assess a company’s machine-learning and computer-vision capability for a leading administrative and financial institution. The optical hardware, the pipeline, and the method were reconstructed blind, to about four-fifths agreement, before the data room opened. Patents, team profiles, and history were read together so earlier technical trouble and later pivots could be forecast. The practice outperformed rival readers by surfacing evidence others had passed over, and co-authored the valuation, the criticism, and the recommendations put in front of senior leadership.

02 Defence

A four-person sensing cell

Computer-vision research and consultation on defence-accelerator demonstrators and on new sensing products. A small team of four. Cross-domain face matching, camera imagery against document photographs, with a true positive rate above 95% on the watch-list task. A white paper on binocular stand-off ranging of people, vehicles, and aircraft by photogrammetry. Classical vision for route optimisation, visual servoing, and resolving pan-tilt and lidar tracks into real-world coordinates. Several edge analytics prototypes with mechanical and electronics specialists. A quarter of the closing written report, including the experimental material, delivered to the defence science sponsor and to external clients.

Overhead view of a thermal camera module, lidar, ribbon cable, and a single-board computer on a wooden bench. — Edge bench. Thermal core, lidar, unbranded compute. From the care and defence hardware work.

03 Care

Thermal models on a weak core

Board-level advice and hands-on computer vision for hyperspectral assistive care. White-label RGB models converted to native thermal inputs. Transformer models for elderly fall detection from abnormal events. Motion detection by a mixture of Gaussians, written to run on low-power edge devices. Models onboarded to AWS Neuron with custom libraries. Architecture for a FLIR Lepton-class thermal system for monitoring elderly and immobile patients. An in-house pair of junior developers mentored through the build.

04 Public interest

Synthetic view, and what to throw away

Consulting across ventures working with local government and with privacy technology. A generative model that produces a stereoscopic view from a single autonomous-vehicle camera, trained on synthetic game footage and commercial simulators. Latent-space traversal so a product can show a user where the model is uncertain, starting from autoencoders. Pre-development of obfuscation networks that keep ethnicity and other demographic features a policy may need, and disconnect the facial traits that identify a specific person.

05 Airframe

A solver too slow to sit inside the loop

Proof of concept on shape optimisation for a commuter aircraft: fuselage and wing, under fuel, engine mass, and passenger count. Familiarisation with SU2, OpenFOAM, and commercial solver practice, then a data-driven turbulence model. Speedups on the order of one to two thousand times against those solvers, at a low and acceptable error. Trained on ground vehicles, with a blind transfer onto subsonic hulls that did not require further training. A mesh representation that hyperparameterises an outline and samples it by regular rays, to cut dimension before the downstream solver. Genetic algorithms and gradient methods, mixing proxy criteria with the real model at a coarser time step. A three-year extension roadmap.

06 Picture

Rotoscope, when the object disappears

AI lead for tools in the visual-effects industry. Panoptic segmentation, vision transformers, feature pyramids, a text encoder, multi-object tracking, U-Nets, captioning models, generative models, and agentic AI, used to improve rotoscope through occlusion. A multi-object tracker deployed so several objects in a shot can be followed when they pass behind one another. Advice on research direction, and on hiring the AI team.

Track reconnection at low frame rate

An on-edge person tracker in the DeepSORT class, extended with a similarity network trained on pairs of human detections, so a broken track can be joined again when the camera is slow.

Clinical attribution in the home

Principal research on a European in-home sensing programme. A lightweight re-identification module, including through apparel change, so wearable medical sensors stay with the right resident. A new RGB-D dataset of about twenty participants, also useful for action recognition.

Fleet parts, made joinable

Interim analysis of corrupted and inconsistent OEM car-part records, so later analytics had something to stand on. A document schema across a document store and object storage.

Agents, retrieval, evaluation

A concentrated build-out of language-model systems: LangChain and LangGraph, Hugging Face, vector stores, prompt work, fine-tuning, and evaluation. End-to-end semantic search, document analysis, and multi-step workflows, with a clear view of context cost and failure.

Papers, then the prototype.

The practice rests on a doctorate in real-world human re-identification, a master’s in cognitive computing, and a first degree in psychology. The doctoral work treated attributes as a mid-level representation: compatible with existing matchers, low-dimensional, and usable as a description when there is no probe image. Later work cut the quadratic annotation burden of camera-pair matching by domain transfer, and asked whether a person can be retrieved by words the model was never trained on.

Publication venues include an oral paper at the British Machine Vision Conference, workshops at the European Conference on Computer Vision and at Computer Vision and Pattern Recognition, an invited chapter on attribute re-identification, and an oral at a UK defence and security doctoral symposium. Topics span attributes in the wild, open-world search from a drone, domain transfer, and persistent multi-target tracking in RGB-D. Postdoctoral work supplied the re-identification module for a European home-sensing programme.

Teaching covered computer vision, data mining, and programming, from support through to new lecture material. Reviewing for IEEE Transactions on Image Processing, Computer Vision and Image Understanding, Neurocomputing, Image and Vision Computing, and related journals. A grant proposal of £130,000 was instigated and co-authored. Early career work also included enterprise web delivery and the IT estate of a small group of companies; that is infrastructure, not the practice’s offer.

Methods in regular use, without a preference for fashion:

PyTorch
TensorFlow
JAX
OpenCV
Hugging Face
LangGraph
ROS
OpenFOAM
SU2
AWS Neuron
Azure

Judgement at the level of the model.

An MVP, or a proof of concept.