Evidence of Work

Evidence of work. Domains, not names.

29 systems across 26 domains. Client names stay with our clients. What remains is the engineering: the problem, what we built, the idea that made it work and what changed.

Vision Speech & voice Language & knowledge Generative media Prediction & decision Sensing & edge Agentic systems

Vision

No. 20 of 29

Pipeline anomaly detection

Oil and gas

The problem: Cracks and leaks across a large pipe network were found by manual inspection — slow, labour-intensive, error-prone and a safety risk.
What we built: A computer-vision system that detects and classifies defects such as cracks and leaks automatically, integrated with existing infrastructure for real-time monitoring and alerts.
The innovation: Deep-learning defect classification on live inspection imagery, turning inspection from periodic manual review into continuous detection.
Outcome: Faster inspection, more accurate detection than manual review, improved safety and lower cost, with real-time alerts enabling preventive action.
Built with: PythonOpenCVTensorFlowKeras

No. 22 of 29

Occupancy and incident detection

Workplace safety · technology sector

The problem: Counting people and spotting accidents across workspaces was manual and unreliable, creating safety and liability risk.
What we built: A vision system that counts people per area, alerts when occupancy limits are breached, and detects incidents such as falls and collisions in real time.
The innovation: Event detection and continuous counting in one video pipeline, with alerts routed straight to the safety team.
Outcome: Instant alerts, better compliance with safety regulations and lower liability.
Built with: OpenCVTensorFlowPyTorch

No. 24 of 29

Vehicle classification for toll supervision

Public sector · transport infrastructure

The problem: Toll supervision relied on manual counts of vehicle numbers and types, producing errors, delays and room for fraud.
What we built: A vision system that counts and classifies vehicles — cars, trucks and buses — at toll plazas in real time, integrated with the toll infrastructure.
The innovation: Real-time detection and classification at the lane, creating an independent record of traffic to supervise collection against.
Outcome: Less manual data collection, faster and more accurate toll data, and detection of toll fraud that increased revenue.
Built with: PythonOpenCVTensorFlowKeras

No. 26 of 29

Crowd proximity and mask detection

Retail · shopping centres

The problem: During the pandemic, a large shopping centre needed to monitor distancing and mask-wearing across busy public spaces — impossible to do manually.
What we built: Real-time computer vision that measured proximity between visitors, highlighted high-risk zones and identified people without masks, alerting staff.
The innovation: Spatial analysis on the centre’s existing camera infrastructure, inferring real-world distances between people from video.
Outcome: Less monitoring effort, faster intervention and restored visitor confidence.
Built with: PythonAzure Spatial Analysis

No. 28 of 29

Parking availability sensing

Hospitality

The problem: Guests had no way of knowing whether parking was available, causing confusion and congestion.
What we built: Computer vision that counts free spaces in real time and publishes availability to a large display board.
The innovation: Occupancy detection from camera views, updated continuously and pushed to signage.
Outcome: Faster parking, a better guest experience and more efficient parking management.

Speech & voice

No. 03 of 29

Voice-to-ticket maintenance requests

Heavy machine manufacturer · fleet operations and maintenance

The problem: Maintenance requests arrived incomplete — missing codes, asset details or context — triggering back-and-forth between operators, dealers and service teams before work could begin.
What we built: A voice assistant inside the client’s fleet operations and maintenance platform. Operators describe the issue in natural speech; the assistant verifies the asset, asks only for what is missing, and writes a complete, structured service ticket with codes, attachments and routing.
The innovation: Speech turned into structured data: free-form speech is mapped onto the ticket schema, and follow-up questions are driven by whichever required fields are still empty.
Outcome: Fewer incomplete requests, less dealer back-and-forth and smoother service workflows.

No. 07 of 29

Real-time meeting translation

Enterprise collaboration

The problem: Multilingual teams and clients lost time and clarity working through translation.
What we built: A meeting add-in that translates speech live — from one-to-one calls to large town halls — with the option of custom voice models.
The innovation: Streaming speech recognition, translation and synthesis chained with latency low enough to keep a conversation natural.
Outcome: Faster, clearer communication across languages and better support for a global customer base.
Built with: Microsoft TeamsAzure Cognitive Services

No. 16 of 29

Call summarisation and topic extraction

Contact centres

The problem: Managers could not keep up with the volume of recorded customer calls.
What we built: Automated summaries of recorded calls, with key phrases and topics extracted for review.
The innovation: Speech recognition and language processing combined into review-ready summaries, turning recordings into a searchable record.
Outcome: Faster review, quicker corrective action and earlier resolution of customer issues.
Built with: Azure Cognitive Services

No. 17 of 29

Call transcription, diarisation and compliance

Contact centres

The problem: Manual transcription and review of customer calls was slow and error-prone, creating quality and compliance risk.
What we built: Automated transcription with speaker diarisation, and analysis for quality and compliance including flagging of prohibited language.
The innovation: Diarised transcripts make every utterance attributable to agent or customer — the foundation for per-agent quality and compliance analysis.
Outcome: Less manual review, real-time alerts, better agent performance and stronger compliance.
Built with: Azure Cognitive Services

No. 27 of 29

Voice-driven multi-cloud monitoring

IT operations · e-commerce

The problem: Monitoring services across several clouds during holiday traffic peaks overwhelmed the operations team.
What we built: Voice skills for smart assistants that report service status and details on request, connected to monitoring across cloud providers.
The innovation: A conversational interface over heterogeneous cloud monitoring: one spoken question, answered across clouds.
Outcome: Faster status checks and less manual monitoring during peak periods.
Built with: Amazon AlexaGoogle HomeCloudWatchGoogle Cloud Monitoring

Language & knowledge

No. 01 of 29

Service diagnostics and knowledge assistant

Heavy machine manufacturer · global after-sales service

The problem: Service teams working on increasingly complex machines had to search manuals, bulletins, historical cases and expert memory to diagnose a fault — slowing every repair and making the outcome depend on who happened to know what.
What we built: A global enterprise application that gives service teams guided diagnostics, troubleshooting paths and instant access to service knowledge, used by thousands of people across the service network.
The innovation: Generative AI grounded in the manufacturer’s own service data: reported symptoms go in; relevant procedures, likely causes and the next diagnostic step come out — turning decades of service knowledge into a working tool rather than a library.
Outcome: In use by thousands of service professionals worldwide, delivering major cost savings and gains in operational efficiency.

No. 04 of 29

On-premise language model for customer support

E-commerce

The problem: Hosted language-model services were expensive, inflexible and required data to leave the company.
What we built: An on-premise language model based on LLaMA, trained with more than 30,000 domain-specific commands and hosted on the client’s own servers.
The innovation: Open-weight model adaptation behind the client’s firewall: full control of data, cost and behaviour.
Outcome: Real-time customer support at lower cost, with data kept on the client’s own servers.
Built with: LLaMAon-premise hosting

No. 05 of 29

Governed knowledge retrieval for field diagnostics

Industrial equipment manufacturing

The problem: Diagnostic knowledge was spread across machinery manuals, historical service tickets and the experience of senior field engineers.
What we built: A governed retrieval and decision-support system: graph-assisted retrieval over technical documentation, on-premise model execution, and source citations surfaced in the technician’s diagnostic workflow.
The innovation: Permission-aware retrieval with on-premise inference, so answers carry citations to authoritative engineering sources and data stays under the client’s control.

No. 06 of 29

Enterprise knowledge repository for language models

Heavy machine manufacturer · enterprise knowledge

The problem: Subject-matter experts, marketing, support and HR each held knowledge in their own formats, with no systematic way to make it usable by large language models.
What we built: A platform for building curated datasets, fine-tuning models and retrieving knowledge through vector search, with a custom embedding infrastructure for data privacy and control.
The innovation: Dataset creation as a self-service workflow for non-specialists, combined with private embeddings and fine-tuning — so domain knowledge improves model behaviour without leaving the organisation’s control.
Outcome: Better model performance on industry-specific tasks, stronger data control, and teams contributing directly to a shared knowledge base.
Built with: Azure OpenAIvector searchcustom embeddings

No. 13 of 29

Generative troubleshooting assistant

Consumer appliances

The problem: Customers with appliance problems faced slow, manual support.
What we built: A fine-tuned GPT-based troubleshooting guide that answers customer questions in natural language, in real time.
The innovation: A language model adapted to the manufacturer’s troubleshooting knowledge and orchestrated with LangChain.
Outcome: Faster resolution, higher customer satisfaction and lower support costs.
Built with: OpenAI GPTLangChain

No. 14 of 29

Invoice extraction with semantic search

Finance operations

The problem: Invoice details were captured by manual scanning and keying.
What we built: Our information-extraction platform, integrated into the client’s systems, using semantic search to pull key details from invoices automatically.
The innovation: Semantic understanding of document content, so key fields are found by meaning across varied invoice layouts.
Outcome: More than 100 documents processed in under a minute, with fewer errors and no manual scanning.
Built with: PythonTensorFlow / PyTorchcloud infrastructure

No. 15 of 29

Insurance document digitisation and summarisation

Insurance

The problem: A large archive of insurance documents was hard to find, read and use.
What we built: Digitisation, summarisation and metadata indexing of insurance documents into a searchable knowledge base for underwriters.
The innovation: Transformer-based summarisation combined with structured metadata indexing, hosted on cloud infrastructure for scale.
Outcome: Faster access to knowledge, fewer manual errors, and underwriters able to draw on the archive when quoting premiums.
Built with: BERTAtlasDBAWS

No. 18 of 29

Discovery document translation

Legal

The problem: A firm working across countries spent significant effort translating discovery documents and comments.
What we built: Automatic language detection and machine translation of discovery documents and comments inside the firm’s document libraries, adapted to each user’s locale.
The innovation: Translation that follows the reader: the detected locale chooses the target language, inside the existing document store.
Outcome: Less translation effort and smoother cross-border collaboration.
Built with: SharePointAzure Cognitive Services

No. 25 of 29

Regional sentiment analysis

Automotive

The problem: A manufacturer could not see how its products and services were perceived from region to region across two countries.
What we built: Real-time sentiment analysis across several social platforms, scored and visualised by region.
The innovation: Sentiment resolved to geography, so perception can be compared market by market.
Outcome: Faster, better-informed product and service decisions.

Generative media

No. 09 of 29

Automated multilingual video dubbing

Media and enterprise content

The problem: Dubbing video into several languages was slow, costly and error-prone.
What we built: A guided pipeline that transcribes, translates, generates new voice tracks and stitches the finished video.
The innovation: Transcription, translation, speech synthesis and alignment orchestrated as a single workflow.
Outcome: Faster, lower-cost localisation for a global audience.
Built with: Azure OpenAIAzure Speech SDK

No. 10 of 29

Enterprise voice generation

Enterprise learning and communications

The problem: Professional voiceover for training and onboarding depended on booking voice talent — and re-recording every time a script changed.
What we built: A platform that turns scripts into lifelike speech in seconds, with selectable voices and tones, multilingual output and ready-to-use audio files.
The innovation: Neural speech synthesis inside an enterprise workflow: controllable tone and pacing, consistent voices across a catalogue, and regeneration instead of re-recording when content changes.
Outcome: Used for training content, onboarding materials and internal communications.

Prediction & decision

No. 19 of 29

Predictive maintenance for aircraft engines

Aviation

The problem: A legacy system could not process terabytes of engine and asset data quickly enough to anticipate failures.
What we built: A predictive-maintenance platform on big-data infrastructure that analyses engine and asset data to detect potential problems early, with 3D visualisation.
The innovation: Distributed processing of terabyte-scale telemetry to surface early failure indicators.
Outcome: Proactive maintenance, improved safety and performance, optimised maintenance schedules and less downtime.
Built with: HadoopHDFSRTableau

No. 23 of 29

Prescriptive spare-part sourcing

Heavy equipment manufacturing · cranes

The problem: Sourcing spare parts for crane assembly was manual, causing delays and extra cost.
What we built: A prescriptive-analytics system that recommends the best sourcing option for each part by weighing cost, availability and delivery time.
The innovation: From prediction to prescription: the system recommends an action, not only a forecast.
Outcome: Faster sourcing, lower part costs and fewer assembly delays.

No. 29 of 29

Race outcome prediction

Sports and racing analytics

The problem: Race forecasting relied on manual analysis of race cards and form — inconsistent and error-prone.
What we built: A predictive system combining computer vision and predictive analytics over race cards, performance history, pedigree, weather and track conditions, with a visual interface for real-time analysis.
The innovation: Multi-source fusion — extracted race-card data, form history and environmental conditions — in one predictive model, processed in the cloud.
Outcome: More accurate, consistent predictions available in real time.

Sensing & edge

No. 11 of 29

Engine Health Monitoring (EHM)

Airline turbine engine manufacturer

The problem: Early signs of engine degradation — abnormal vibration, shifts in acoustic signature — are subtle, buried in operating noise, and often recognised only after they have grown into maintenance events.
What we built: An engine health monitoring system that captures acoustic and vibration signals, processes them to isolate fault signatures, and flags anomalies for engineering and maintenance teams.
The innovation: Digital signal processing at the core: spectral analysis separates the acoustic and vibration signatures of developing faults from normal operating noise, and models classify the likely condition from those signatures.
Outcome: Earlier detection of developing faults, and a faster path from signal to maintenance decision.

No. 12 of 29

Acoustic anomaly detection on CNC spindle lines

Automotive component manufacturing

The problem: Intermittent bearing wear in high-speed spindles led to tool fractures that damaged parts before visual inspection could catch them.
What we built: An on-device acoustic model on industrial edge hardware, isolating harmonic anomalies from background plant vibration.
The innovation: A quantised model sized for edge accelerators, with inference entirely on-device and no cloud dependency.

No. 21 of 29

Equipment telematics and fleet intelligence

Heavy and off-road equipment · fleet management

The problem: Fleet owners could not see malfunctions or violations — geofence breaches, overloads, fuel pilferage — until after the damage was done.
What we built: A cloud platform that collects sensor and CAN-bus data from connected machines, with equipment-specific digital capacity charts for overload detection, geofencing, fuel-pilferage detection and business-intelligence reporting, offered as a subscription service.
The innovation: Digital capacity charts for each machine model turn raw telemetry into overload violations in real time, while in-equipment devices keep data flowing over constrained cellular links.
Outcome: Real-time asset management, proactive detection of violations, lower operating costs, and better utilisation and safety.
Built with: Microsoft stackMQTTCAN buscellular telemetry

Agentic systems

No. 02 of 29

Agentic certification audits in an isolated network

Aerospace and defence

The problem: Certification audits required cross-checking telemetry manifests, material certificates and test logs — manually, over weeks.
What we built: An OpenWolfPack deployment in an isolated on-premise cluster, with checkpointed execution and tamper-evident audit logging.
The innovation: Deterministic checkpoints make audits resumable, with human sign-off at every certification boundary and execution traces reviewable by compliance officers.

Every record is anonymised by design. Ask about a comparable problem and we will tell you what we can.

Discuss a comparable problem.

Start a conversation See the evidence gallery