Services
Four disciplines, engaged separately or in combination.
Each engagement begins from the constraints (the data, its required location, the users, and the governance under which the resulting system will operate), and those constraints, rather than a preferred architecture or vendor, shape the work that follows. The four disciplines below describe the shapes that engagement most commonly takes.
AI product engineering
Full-stack delivery of language-model platforms, agent frameworks, and bespoke artificial-intelligence interfaces, all deployed inside the client's tenant, with a data architecture designed to keep it there.
- Language-model applications across OpenAI, Anthropic, Azure OpenAI, and self-hosted models, including extensions of open-source platforms such as LibreChat.
- Agent and tool-use architectures, including Model Context Protocol (MCP) servers for enterprise integration with Microsoft 365 and SharePoint.
- Custom Copilot-style extensions, Microsoft Teams applications, and internal Copilot workflows.
- Evaluation harnesses, prompt-regression testing, and model-routing strategy for production multi-model deployments.
- Targeted fine-tuning and lightweight post-training, applied where the evaluation demonstrates that it is justified.
Reference engagement
Chat:R, an extended LibreChat deployment serving approximately one thousand staff at Rud Pedersen Group, with a purpose-built SharePoint Model Context Protocol gateway and three bespoke enterprise features.
Read the case study →Applied machine learning
Computer vision, time-series forecasting, and interpretable predictive modelling for clinical, operational, and financial settings. Methods are chosen by problem rather than by trend; baselines are built where practical; honest evaluation precedes delivery.
- Computer vision and segmentation on real imagery, including deep-learning and classical baselines for comparison.
- Time-series forecasting and continuous-learning systems, typically with gradient-boosted methods and careful validation.
- Interpretable predictive modelling, with SHAP and related methods, for settings in which the explanation is part of the deliverable.
- Retrieval-augmented generation pipelines over document and briefing corpora: ingestion, chunking, embedding, hybrid retrieval, reranking, and evaluation.
- Public-dataset joins (HM Land Registry, Energy Performance Certificate, Office for National Statistics) and geospatial analytics where the analysis calls for it.
Prior work
A U-Net cell-spheroid segmentation model for a UCL Royal Free Hospital research group, with a validation intersection-over-union of 0.968 and Dice coefficient of 0.983; residual-value forecasting across an electric-vehicle fleet (gradient-boosted trees with continuous learning); an interpretable house-price model with SHAP-based explanation over public-register data.
Background on the founder →Platform and deployment
Tenant-bound cloud deployment of AI systems: the infrastructure, identity, and operational choices that decide whether a working prototype can be adopted in production. The studio works at an architecturally literate level here, and pairs with specialist cloud engineers when the engagement calls for deep individual-contributor platform work.
- Azure-first service selection: App Service, Virtual Machines, Container Instances, Storage, Key Vault, AI Search.
- Infrastructure-as-code with Bicep; GitHub Actions with federated OpenID Connect authentication; multi-environment continuous integration and deployment.
- Identity and authorisation design: Microsoft Entra single sign-on, domain-scoped access, managed identities, private endpoints.
- Observability through Application Insights and Log Analytics; incident response and runbook authorship.
- Multi-cloud capable (Amazon Web Services, Google Cloud Platform) where the engagement constraints prefer it; architecture is led by the problem rather than by a single vendor preference.
Reference engagement
The Chat:R deployment operates across three Azure environments, with Bicep-authored infrastructure, federated GitHub Actions pipelines, Microsoft Entra single sign-on, and a considered pivot from managed Kubernetes to virtual machines that reduced monthly operating cost by approximately one half.
See the infrastructure notes →AI strategy and governance
Published advisory, programme design, and readiness assessment for organisations moving past proof-of-concept artificial intelligence. Informed by peer-reviewed and government-published work, including the founder's own contributions.
- AI risk, governance, and cyber-security advisory, drawing directly on the UK Government DSIT whitepaper on cyber-security risks to artificial intelligence (Barua et al., 2024).
- Sector-specific deployment guidance, drawing on the Jersey Finance published guide to artificial intelligence in Jersey's finance industry (McCay and Barua, 2024).
- Generative-AI programme design at enterprise scale, including portfolio planning, use-case triage, and cross-divisional delivery.
- Build-versus-buy assessments across enterprise platforms (ChatGPT Enterprise, Microsoft Copilot, self-hosted, and bespoke), presented with honest cost and capability modelling.
- Internal enablement: workshops, written training material, and champions-network design for organisations adopting AI at scale.
Prior work
A generative-AI programme at Grant Thornton of approximately ten million pounds in committed value, comprising twenty-five-plus production use cases, eight million pounds-plus in projected efficiency savings, and two thousand-plus staff demonstrations. Named co-authorship on two 2024 publications: the UK Government DSIT whitepaper on cyber-security risks to artificial intelligence, and the Jersey Finance guide to artificial intelligence in Jersey's finance industry.
Background on the founder →