Isaac Griffiths

I am Isaac Griffiths, an AI Engineer in GSK R&D Tech and the technical lead of an agentic product, one of the first genuinely agentic systems in this part of the business. The tools I build are internal, which means they have to survive real users, real compliance and a real queue. I joined as a degree apprentice in 2022. In August 2026 I was promoted to manager grade.

It started with a dull observation. Generic request forms let people submit almost nothing, so an engineer's first reply is spent asking for what should have arrived with the ticket. I built a triage layer end to end, as the only person who wrote code for it, that asks context-aware follow-up questions at intake, so the ticket arrives actionable instead of arriving twice. LangGraph, FastAPI and Next.js on Google Cloud Run, an MCP tool server on the live ServiceNow API, and the pipeline that ships it, all mine.

The way I work is closer to forward deployed than to a backlog. I sit with the team that has the problem and build alongside them: something working on screen inside the session, then reshaped around what they actually do with it. I build with Claude Code as my primary tool. It writes most of the Python and TypeScript; I specify the system, direct it, then read, review, test and debug whatever comes out. That second half decides whether anything ships. About 70% of my work is backend and cloud, 30% front end.

Before that I spent three years as a lab IT engineer, keeping regulated scientific environments running: incident response, decommissioning, OT network segmentation and a lot of instruments that do not tolerate downtime. Three years of that teaches you something. The interesting part of a system is what it does on a bad day.

• Aug 2026 – present

  AI EngineerCurrent

  GSK R&D Tech (Pharmaceuticals) · Stevenage and London

  Technical lead of an agentic product in Scientific Data Engineering, on the Data and Analytics Engineering Manager job profile. I own the architecture, the delivery plan and the route from working prototype through architecture review to production, and I run the Google Cloud side end to end: Cloud Run, the GitHub Actions pipeline into Artifact Registry, Secret Manager and Cloud Monitoring. About 70% of my time goes on the backend and cloud, 30% on the front end. I communicate with principal data scientists, lab researchers and IT compliance, bring them to the same architecture approvals, and mentor apprentices on prompt evaluation, structured outputs and secure API design.

  • Python
  • FastAPI
  • LangGraph
  • Pydantic
  • Azure OpenAI
  • Gemini
  • Google Cloud Run
  • Artifact Registry
  • GitHub Actions
  • Claude Code
• Sep 2025 – Aug 2026

  AI Engineer, Degree Apprentice

  GSK R&D Tech (Pharmaceuticals) · Stevenage

  I built the intake and triage agent end to end, as the only person who wrote code for it: it turns incomplete lab support requests into tickets an engineer can act on. LangGraph, FastAPI and Next.js on Google Cloud Run, an MCP tool server on the live ServiceNow API, and the infrastructure as code through GitHub Actions and Artifact Registry. I co-authored and presented the architecture, then ran the testing cycle: 50+ structured cases including adversarial inputs, pytest and Jest in CI, Postman suites, trace-identifier logging across agent boundaries and peer review on every change. I rebuilt a compliant replacement for an internal application whose source code had been lost, verifying its behaviour with the scientists who use it, and prototyped an agentic instrument booking flow that I am improving with those same scientists.

  • Python
  • LangGraph
  • MCP
  • FastAPI
  • Next.js
  • TypeScript
  • Docker
  • Google Cloud Run
  • Artifact Registry
  • GitHub Actions
  • pytest
  • SonarQube
• Sep 2022 – Sep 2025

  Laboratory IT Engineer, Degree Apprentice

  GSK R&D (Pharmaceuticals) · Stevenage

  I kept regulated laboratory environments running across specialist instruments, endpoints and OT networks. When CrowdStrike took the site down I was the first at GSK Stevenage to restore an affected lab PC, and the BitLocker approach I worked out became the site's remediation procedure: 250+ systems back inside two days. I led the removal of 50+ Windows 7 systems under the GSK Legacy Programme, automated the lab PC build, supported 250+ specialist lab devices and worked with cybersecurity on OT network segmentation. Eleven GSK recognitions in three years: one Gold, one Silver, five Bronze and four peer awards.

  • Windows
  • PowerShell
  • WDS/MDT
  • Active Directory
  • BitLocker
  • VLANs and firewall policy
  • OT segmentation
  • ServiceNow

Languages

Python, TypeScript, JavaScript, SQL, HTML, CSSPython on the GSK triage and booking backends; TypeScript across the triage UI, the legacy rebuild, TwinList, ShipStamp and this site.

AI & agentic

LangChain, LangGraph, Prompt engineering, Structured outputs, Tool calling, OpenAI, Anthropic, GeminiLangGraph state machines on the GSK triage agent (2025) and booking agent (2026), with an MCP server into ServiceNow. GPT-4o via Azure and Gemini at GSK; Claude in TwinList.

Web & APIs

Next.js, React, FastAPI, PostgreSQL, REST APIs, Electron, Integration designFastAPI behind the triage and booking agents; Next.js on the triage UI, booking dashboards and the legacy rebuild; PostgreSQL with pgvector and Electron in TwinList and ShipStamp; Swift and SwiftUI in SMBDrop.

Working with stakeholders

Embedded delivery, Pair-building, Requirements by demonstration, User testing, Adoption & handoverEvery GSK build since 2025: requirements settled with scientists in the room, then UAT, sign-off and training across lab teams.

Automation

n8n, Workflow orchestration, Process automation, ServiceNowServiceNow through the triage agent at GSK; n8n at home for the AI news digest that feeds my content.

Cloud & delivery

Google Cloud Platform, Microsoft Azure, GitHub, Cloudflare, DockerTriage agent in Docker on Cloud Run via GitHub Actions and Artifact Registry, all of it mine; Azure OpenAI through GSK's Kong gateway. Cloudflare Workers, D1, Pages and R2 for licensing and this site.

AI-native development

Claude Code, GitHub Copilot agent, Cursor, Codex, Ollama, Open WebUI, Rapid prototyping, Review, testing and debugging of generated codeAt GSK: Cursor at the start in 2025, then the Copilot coding agent, now Claude Code only. At home: Claude Code, Cursor, Codex and a self-hosted model on an old gaming GPU. Claude Code writes most of the code. I specify, review, test and debug it: pytest and Jest in CI, Postman, SonarQube, Playwright end to end on the products.

Infrastructure

Windows & Linux, Networking & VLANs, Active Directory, WDS/MDT imaging, PowerShell, OT segmentation, Proxmox, UnraidThree years in GSK's Stevenage labs: PC build automation, instrument networking, OT segmentation, the CrowdStrike recovery. Proxmox and Unraid at home.

Things I have used to ship something real, not things I have read about.

• Personal project

  ShipStamp

  A Windows desktop app in Electron, React and TypeScript, built end to end. It pulls courier label PDFs from a Gmail or Microsoft inbox, fingerprints each layout to discover new courier variants on its own, and prints straight to a thermal label printer. Licensing and entitlements run on a Cloudflare Worker with D1 and RevenueCat.

  Built for my own reselling first. The site and waitlist are live at shipstamp.co.uk; mailbox access is in review with Google and Microsoft.

  • Electron
  • React
  • TypeScript
  • Gmail API
  • Microsoft Graph
  • Cloudflare Workers + D1
  • End to end
• Personal project

  TwinList

  A Next.js and Electron app, built end to end for my own reselling and used on real listings. Product photos go in, vision models draft the listing details, I approve them in a review deck, and a Chrome extension posts them to the live marketplace forms, escalating any field it cannot fill with confidence. Corrections feed a pgvector example bank, so it learns my judgement.

  Photos become posted listings, with me editing rather than typing.

  • Next.js
  • Electron
  • Vercel AI SDK
  • PostgreSQL + pgvector
  • Chrome extension
  • eBay Sell API
• Personal project · open source

  SMBDrop

  A free, open-source iOS app that sends photos, videos and files from an iPhone to an SMB share and pulls them back. Built end to end from a Windows machine: SwiftUI app, Share Extension, Keychain, and a fastlane pipeline on GitHub-hosted macOS runners to TestFlight.

  Public on GitHub. App Store release in progress.

  • Swift
  • SwiftUI
  • SMB
  • Share Extension
  • fastlane
  • GitHub Actions

• 2022 – 2026

  BSc (Hons) Digital and Technology Solutions (Software Development)

  Degree apprenticeship · Anglia Ruskin University, Peterborough

  Completed alongside full-time delivery at GSK.

  First Class Honours · Awarded 2 September 2026. Overall 73%, eight modules at First.

  Selected results

  Programming and Networking IoT Project

  83%

  Computer and Network Technologies

  83%

  Information Security Management and Governance

  81%

  Object Orientated Modelling and Data Design

  76%

  Enterprise Analytics

  75%

  BTEC ICT: Distinction* (Claires Court School, 2022). A Level Geography: B (Claires Court School, 2022). 10 GCSEs, grades 5–8 (including Maths (8) and ICT (8)).