~/cv
Rob Gilks
profile
I build AI products end-to-end: fine-tuning transformer models, deploying them on serverless GPUs and the edge, and shipping polished web apps around them. At Cambridge University Press & Assessment I led the team behind the APIs powering AI auto-marking of written and spoken English exams - a hands-on lead, top committer on the core services while running the platform to 99.98% uptime.
Documentation-first: every project ships with architecture docs, decision records and diagrams kept as code. There's a whole page on that - how I run a team.
skills
ai & ml
web & cloud
languages
platform & ops
practice
experience
Independent software studio and consultancy - full-time focus since May 2026.
- Design, build and run a portfolio of live AI products for language learning and assessment (below), deployed on Cloudflare's edge with serverless GPU inference on Modal.
- Fine-tune and deploy transformer models: CEFR essay scoring (DeBERTa-v3 - 99.6% adjacent accuracy), grammatical error correction (GECToR, seq2seq, LanguageTool) and on-device speech recognition (Whisper).
- Documentation and testing as a habit: 20+ repos share a house style - ARCHITECTURE and SPEC docs, decision records, and Graphviz/Mermaid diagrams rendered and validated in CI - with Playwright/Vitest suites throughout.
- Author of five production AI/ML engineering training courses, from serverless edge architecture to fine-tuning transformers and ML deployment.
- Led the Assessment & Research Capabilities (ARC) API team - building and operating the Text and Speech APIs that power automated marking across Cambridge assessments, including Write & Improve, Speak & Improve and Linguaskill.
- A hands-on lead: top committer on the Text API and its orchestration service (Clojure event-driven microservices on AWS) and a lead contributor to the five-environment Terraform estate.
- Ran the platform to 99.98% uptime over 12 months, with Prometheus/Grafana observability, PagerDuty incident response and DORA delivery metrics (~1% change failure rate).
- Took AI models from research hand-off to production: transformer and XGBoost automarkers, off-topic and anomaly detectors, and Whisper-based speech recognition with speaker verification for malpractice detection; evaluated marking quality with Claude on AWS Bedrock.
- Built the team's documentation culture - architecture, RFCs, runbooks, incident response, onboarding and a public API documentation site.
- Senior engineer on Write & Improve and Speak & Improve - AI feedback tools that help English learners worldwide, built in Clojure around custom marking models. The company was merged into Cambridge in 2025.
- Tech lead on a system enabling clinical research sites to communicate with trial patients via SMS.
- Envisioned a customisable trading platform for brokerages with the founder, built the early prototypes, then built up the team and engineering practices.
- Set the technical direction for a precious-metals derivatives trading platform and transitioned the development team from London to Ukraine.
- Rebuilt a platform for running collaborative role-play training experiences online, alongside other engagements.
- Ten years building interactive products and campaigns for Nike, Heineken, Nissan, Fiat, the BBC, Grey Goose and Ford - including Nike Pro Genius, Heineken Star Player and the BBC's Story of Life.
- LTA - TotalTennis club-website platform and live tournament scoring (Nottingham, Eastbourne). Oxford University LTG - Java learning-tools middleware. Play Sport / Pretzel Logic - sports-club and government web systems. Ampac Australia - fire-detection systems, from electronics technician to R&D engineer.
selected projects
comprehendo
Adaptive reading coach in 16 languages - AI-generated passages at CEFR level, tap-to-translate, TTS. In live testing with an early user cohort.
talata
AI essay scoring and feedback - a fine-tuned DeBERTa-v3 multi-trait scorer and dual grammar-correction models, served as five GPU microservices on Modal.
speako
Pronunciation coach running entirely in-browser - local Whisper plus a fine-tuned DeBERTa CEFR model via Transformers.js and WebGPU. Nothing leaves the device.
galacto
Black-hole accretion simulation - 130k+ particles on GPU compute shaders, with sound synthesised from the galaxy's motion.
rowspire
Strategy game with two in-browser AI opponents - a bitboard minimax solver and an AlphaZero-style neural-network MCTS, in Rust/WASM.
Full portfolio on the home page - everything above is live, documented and tested.
education
BE Electronic Engineering - The University of Western Australia.
The paper version
Same content, two tidy pages, ready for your ATS of choice.
last updated: 2026-07