Agentic AI, retrieval-augmented generation (RAG), and LLM-powered automation, built in Python. Four years of data experience inside a regulated insurer, now applied to building AI tools for real business problems.
In plain terms: I build software that uses AI to answer questions, automate repetitive work, and make decisions — the kind of tools a business uses to save time and cut cost. Every project below links to the working code.
Each card explains what the project is, what it is used for, where it applies in industry, and links to the code. No technical background needed.
An AI system that answers questions about a document by finding the right passages first, then writing a grounded answer — and it decides on its own whether to search the document, search the web, or answer directly.
Letting a person ask plain questions of a large document (a report, contract, or manual) and get accurate, sourced answers instead of reading the whole thing.
I did not just build it — I measured it. Using an industry evaluation tool (RAGAS), the answers scored 97% on faithfulness (staying true to the source).
✓ 0.97 faithfulness · 0.92 precisionInsurance & finance: policy wordings, claim files, regulatory rulebooks. Pharmaceutical & medical: trial protocols, drug monographs and clinical guidelines, as a research aid with a qualified person reviewing every answer. Government & defence: legislation, procurement papers, technical manuals. Environment: long scientific and impact reports. Because answers are scored for faithfulness to the source, it suits regulated settings where an unsupported answer carries real risk.
A chat-based research assistant that calls live tools — web search, competitor analysis, pricing and news — through a custom tool server I built, then packages the results back into the conversation.
Turning a slow manual research task into something a non-technical person can do by asking in a chat window, with the AI fetching and structuring the information.
The tool server is packaged with Docker, so it runs identically on any machine — a step toward production-ready, not just a demo.
✓ Containerised & runnableFinance & insurance: competitor monitoring, supplier and counterparty due diligence. Pharmaceutical: tracking rival pipelines and public regulatory announcements. Government & defence: open-source information gathering from public sources. Environment: monitoring published data on emissions, incidents and policy change. Any team that currently pays people to read the internet and summarise it.
A team of four AI agents — planner, researcher, analyst and strategist — working in sequence to produce a market-entry strategy, built two different ways so the approaches can be compared directly.
Automating a piece of business analysis that would normally take a team of people, while keeping every claim traceable to its source.
Every fact the agents use is tagged with where it came from, and the whole run is scored on quality — so the output is auditable, not a black box.
Pharmaceutical: market-entry planning for a new product or territory. Finance & insurance: product launch, pricing and competitor research. Government: policy option appraisal, where each recommendation must be traceable to evidence. Research: structured evidence reviews in which every claim carries its source. The audit trail is the point — it is built for settings where you must show your working.
A group of AI agents that discuss incoming support tickets among themselves and decide, together, which ones to prioritise.
Helping a team handle a flood of requests by having AI sort and rank them first, so people focus on what matters most.
The agents coordinate with each other (agent-to-agent), rather than a single AI making every call alone.
Insurance: first-notification-of-loss and claims triage, routing and ranking incoming claims. Medical: sorting patient enquiries and administrative requests so clinical staff see urgent items first — ranking only, never a clinical decision. Government: citizen service requests. Defence & environment: incident and fault reports ranked by severity. Any high-volume queue where the cost of missing the urgent item is high.
A set of AI agents (research, funding, pitch coach, critic) that plan an organisation's next steps, share a common memory, and improve their own output through a review-and-refine loop.
Producing a first-draft plan — market research, funding options and a pitch outline — that a person then refines rather than starting from a blank page.
It runs with no paid AI service by default (built-in offline mode), so it can be demonstrated for free, and a critic agent checks the work before it is finalised.
Government & public sector: drafting grant and innovation-funding applications. Environment: planning and costing sustainability initiatives. Pharmaceutical & medical: early-stage research spin-out and translation planning. Finance: investment screening, where the critic agent stress-tests the case before a person reads it. The offline mode also suits organisations that cannot send data to an external AI service.
A tool server that gives any AI agent live capabilities — company overview, competitor list, product portfolio, pricing snapshot and news pulse — using MCP, the emerging open standard for connecting AI to external systems.
Letting an AI assistant reach real, current information from outside itself, instead of relying only on what it learned during training.
It is packaged with Docker and genuinely runs: the container serves the tools over the network, with credentials injected at runtime rather than written into the code.
✓ Docker · FastMCP · 5 live toolsFinance & insurance: due diligence and competitor monitoring. Pharmaceutical: pipeline and regulatory tracking. Government & defence: open-source information gathering. Environment: monitoring published data. The pattern matters more than these particular tools — any internal system can be exposed to an AI agent this way, which is how most organisations will connect AI to their own data.
I am an AI engineer specialising in agentic systems — I build and evaluate LLM-powered agents, RAG pipelines and automation in Python. What makes my work different is where I have come from: four years inside a regulated (FCA) insurer, working hands-on with claims and risk data in SQL and Python. That means I build AI for real business problems — cost, accuracy and auditability — not just demos.
I care about whether things actually work. Across my projects you will see evaluation, measurement and observability built in, because "it produced an answer" and "it produced the right answer" are different things. I would rather find a failure in evaluation than in production.
I am commencing an MSc in Artificial Intelligence at the University of Hull in 2026, and I use AI tools (Claude, Cursor, Copilot) daily in my own work. I am a British citizen based in the UK, open to remote and hybrid roles and willing to relocate internationally.
Open to AI Engineer, Applied AI and Agentic AI roles — remote or hybrid.
chait6982@gmail.com