Skip to content

I put AI agents into production for law firms and wealth managers.

I'm a founding Forward Deployed Engineer at Palindrome, where I build the AI agents that draft advice reports for UK wealth managers. I work with each client firm from scoping to production: learning how its advisers write, building the system that drafts the report, and testing it with evaluations before each release.

LinkedIn (opens in a new tab)GitHub (opens in a new tab)

Portrait photo

About

Before that I was a data scientist at Fletchers Solicitors, a law firm, where I put LLM agents into production.

I started in biology. I studied Biological Sciences at Imperial College London and graduated in 2021 with a First, having made the Dean’s List in my first and second years.

I then moved into computing and went back to Imperial for an MSc in Computer Science, which I finished with a Distinction in 2024. I’m co-first author of ForestVO, published in IEEE Robotics and Automation Letters in 2025. Visual odometry works out how a camera moves from the images it records. Our method keeps baseline accuracy while using 25% of the keypoints.

Plot of the camera path estimated by ForestVO against the ground truth path, in metres; the two lines stay close along the whole route.
The camera path ForestVO estimated, against the ground truth, in metres. © 2025 IEEE. From Pritchard et al., ForestVO, IEEE RA-L 2025.

Work

  1. Founding Forward Deployed Engineer

    Apr 2026 – Present

    AI platform for UK wealth managers, London

    1. AI advice reports in production for wealth managers

      1 hourfor a report that took a week

      Problem. Advisers at UK wealth managers wrote each client's advice report by hand, and a complex report took about a week.

      What I built. Shipped 5 AI advice-report products to production in 3 months for two wealth managers across 4 countries. AI agents draft each report from the client's records, and the adviser reviews it.

    2. Cutting company-wide LLM costs with prompt caching

      67%off the LLM cost of each report

      Problem. The platform's most expensive LLM calls, for document generation, repeated the same instructions and client data on each request but were not getting cache hits.

      What I built. Reordered the prompts so shared instructions and client data come first and are cached, then rolled the change out across document generation, compliance checks, the chatbot and AI editing.

    3. An evaluation system that gates each release

      Problem. Palindrome's clients are regulated wealth managers, so a regression in an AI advice report has to be caught before it reaches them.

      What I built. Built an evaluation system of synthetic client cases, scored by rule checks and an LLM judge, that gates each production release. The company refined it in workshops with the FCA as part of the regulator's AI Live Testing programme.

      Result. Stopped AI report regressions reaching clients.

      Synthetic client cases go to the agents, which draft each report. Rule checks and an LLM judge score every draft. At the release gate, a pass goes to production release; a fail stops the regression before it reaches clients.Syntheticclient casesAgents drafteach reportRule checksLLM judgeRelease gatePassProductionreleaseFailRegressionstoppedSynthetic client cases go to the agents, which draft each report. Rule checks and an LLM judge score every draft. At the release gate, a pass goes to production release; a fail stops the regression before it reaches clients.Syntheticclient casesAgents drafteach reportRule checksLLM judgeRelease gateFailRegressionstoppedPassProductionrelease
  2. Fletchers Solicitors

    Data Scientist

    Oct 2024 – Apr 2026

    UK personal-injury and medical-negligence law firm, London

    1. A WhatsApp agent that wins back written-off leads

      10%of written-off leads won back

      Problem. A UK personal-injury and medical-negligence law firm wrote off leads it had failed to reach by phone.

      What I built. Built a production WhatsApp agent in PydanticAI that answers each lead's questions from the firm's knowledge base and books them a call in Salesforce. A backend engineer built the Salesforce connection.

  3. Stark Software Group

    Asset Analyst

    May 2023 – Sep 2023

    London

    • Automated the monthly asset report on my own initiative, cutting it from 4 hours to 2 minutes with an automated Excel dashboard

Outside work

Education and research

Imperial College London

2024

MSc Computer Science (Distinction)

Imperial College London

2021

BSc Biological Sciences (First Class Honours)Dean's List, Years 1 and 2

IEEE Robotics and Automation Letters

2025

ForestVO: Enhancing Visual Odometry in Forest Environments through ForestGlue (opens in a new tab)Co-first author. Baseline accuracy with 25% of keypoints.

Skills

LLMs and agents
Multi-agent systems, PydanticAI, LangChain, Claude Code, Codex, RAG, Evals, Regression testing, LLM-as-judge, Guardrails, Human-in-the-loop, Prompt caching
Engineering
Python, SQL, FastAPI, Azure (Functions, OpenAI, Speech), Docker, PostgreSQL, PyTorch