Callum Cummins

bet365 · Manchester, UK

• Solely designed and built the Next.js/TypeScript frontend for a compliance investigation platform within three months of joining, taking it from an empty repo to production across multiple regions. It cut case lookup from hours to ~8 minutes, and the Python / Azure OpenAI reporting workflow I co-developed alongside it gave analysts 10x+ throughput.
• Own features end to end on a .NET 8 / React trading operations platform, from the stored procedures and API endpoints up to the UI, state and a 26-case test suite.
• Tracked down why a high-traffic events API was slow and found several separate causes, including Redis cache stampedes, a missing index, a ~5s DataTable step and serializer settings being rebuilt ~111k times per request, and diagnosed connection pool exhaustion along the way.
• Fixed a save bug that turned out to have five separate causes, from a string comparison to a stale client cache and a discarded PUT response, then moved the business rule server-side so it can't drift again.
• Built audit logging for a regulatory compliance platform, capturing what changed at write time with an EF Core interceptor, plus a reusable incremental sync from Postgres to BigQuery running on Hangfire.
• Reviewed the architecture of an LLM-based federated search proof of concept before it was built, flagging prompt injection, token storage and an undesigned streaming path.

Keele University · First Class Honours

• Key modules: Software Development Management, Advanced Databases and Applications, Communications and Networks, Computational and Artificial Intelligence
• Dissertation: First Class grade of 80%. Gym Goer Web App: Developed a full-stack fitness tracking application with predictive ML capabilities, achieving 4.2/5 user satisfaction and 90% recommendation rate.

Coleg Cambria · Merit, Merit, Distinction (MMD)

• Key modules: Computer Systems, Computer Networks, Cyber Security, 3D Animation, Game Design, IT Project Management

Stack: Python, Semantic Kernel, Ollama, OpenAI API, PySide6 (Qt), faster-whisper, openWakeWord, Kokoro

• A voice assistant I talk to like someone in the room, built on Microsoft Semantic Kernel. It runs on local Ollama or cloud OpenAI, switchable mid-conversation by voice, and uses function-calling plugins to control my PC, read my screen, search the web, manage notes and generate images. It also remembers past conversations through embedding-based semantic recall.
• Got warm reply latency down from ~56s to under 1s by turning off reasoning for voice, keeping the model resident in VRAM and streaming each sentence into TTS as soon as it's finished. The voice pipeline (wake word, Whisper, agent, Kokoro) runs as a Qt state machine and is covered by 100+ pytest tests, ruff and mypy.

Stack: C# / ASP.NET Core (.NET 10), EF Core, Supabase Postgres, React 19, TypeScript, Vite, Cloudflare, Docker

• Built the whole site for a small clothing brand on my own: a C# / ASP.NET Core API on Supabase Postgres and a React/TypeScript storefront on Cloudflare, with every product, photo and banner coming from the database.
• Gave the owner her own admin panel on a subdomain where she adds new pieces, uploads photos, bulk-edits items and swaps the homepage banner on the fly, all behind JWT-protected endpoints and with no code changes or redeploys.

Stack: Python, PyTorch, BERT, Tkinter, Pandas, Hugging Face

• Fine-tuned BERT with PyTorch on CUDA GPUs to spot phishing emails, reaching 97.7% validation accuracy (0.98 precision, 0.97 recall) on 3.7K+ samples.
• Wrapped it in a simple Tkinter app: paste an email or load it from an image via OCR, and get a verdict in real time.

Stack: Next.js, TypeScript, Three.js, React Three Fiber, D3-Geo, Zustand, SWR, Google Maps API, OpenAI API, Tailwind CSS, Jest

• An interactive 3D globe of 190+ countries built with Three.js and D3-Geo. Select a country and get an AI-generated summary from the OpenAI API in under 4 seconds.