Katie Hemus
building the infrastructure that makes AI work in production
"I built a RAG system before RAG was a word. An LLM gateway before they were a product category. Twelve years of finding where the puck is going."
Senior platform & infrastructure engineer, operating at Staff level. Previously built the AI backbone for PortSwigger's flagship product. Now leading platform transformation at Civica.
Get in touch → See the work
How I work
Discover
On-site or remote-embedded. Map the actual environment - auth, networks, data shapes, the undocumented edges - before touching code.
Prototype
Working slice in days, not a quarter. Smallest proof the integration runs in your world. Fast feedback beats a perfect plan.
Harden
Production-real - error paths, observability, retries, secrets handling, the boring things that decide whether it survives real traffic.
Handoff
Docs your team can use, a walkthrough, and a clean exit. The goal is your engineers owning it - not a dependency on me.
Platform & infrastructure
LLM Gateways, RAG pipelines, MCP servers, agent loops, closed-source & HuggingFace models, pgvector + PostgreSQL, LiteLLM, Qdrant, Pinecone, embedding models, agent infrastructure - the full stack from model to production. Early AWS SageMaker adopter. Discovery work with TensorFlow, PyTorch, Jupyter.
EKS, Helm (chart authoring, push deploys), ArgoCD, Kargo, CRDs, XRD Composition, Karpenter, Cilium/Hubble (eBPF), OPA/Gatekeeper, Kyverno. Built and operated three production K8s clusters.
Terraform, AWS, GCP. Infrastructure-as-code for platform services, CI/CD pipelines, and AI workloads.
Supply chain security, multi-tenant isolation, sandboxed execution, least-privilege IAM, SBOMs. Security engineering across every role: threat detection, compliance audits, cluster hardening. WIZ, GuardDuty, CloudTrail, Semgrep, Trivy, ClamAV, OpenSSL, AWS KMS.
Nomic Atlas + Nomic Embed for AI-native data mapping. Postgres + pgvector brain servers. Semantic clustering, embedding-space navigation, knowledge graph construction. Obsidian for structured knowledge management.
Vercel for rapid prototyping - ship a working demo in hours. Caddy + Let's Encrypt for production reverse-proxy and TLS on self-hosted infrastructure. Multi-environment: prototype fast on Vercel, production-grade on bare metal.
How I got here
The foundation
Infrastructure engineer across healthtech, media, and ecommerce. Selected for Co-op's New Ventures team. Built an online pharmacy product (AWS, Kubernetes, NHS Spine/EMIS/TPP) that was subsequently acquired. Renewed every 6 months for 4 years with increasing rates. Then ITV and Booking.com: Love Island catch-up at ITV (video on demand to millions, bursty traffic) and Kubernetes migrations at broadcast and global-traffic scale. 12-factor, observability-first, everything in git. Boring fundamentals, drilled in across enough industries that they stopped being opinions.
Foresight
Built a RAG system for engineering docs - vector embeddings, semantic search - pre-ChatGPT-3, before RAG was a product category. Tech support adopted it for customer FAQs. Gave a company-wide talk on vector databases. Read foundational papers including Attention Is All You Need. Earned a seat on the AI research team. That pattern again: built the thing before it had a name.
Proof
Built scan template generation for vulnerability scans - describe what a scan needs to do, get a working template. AWS Lambda PoC → Kubernetes → shipped in the flagship product. Wired in PromptLayer for prompt observability before teams knew they needed it. Steady stream of AI prototypes figuring out what this stuff was actually good for.
Scale
Architected the LLM Gateway - the AI backbone for a flagship product. LiteLLM routing, EKS, Helm, Prometheus/Loki. Built 3 K8s clusters. Deploy cycles 2 weeks → minutes. CI/CD runtimes cut 95%. MCP servers for GitHub and Jira. Worked with Ollie Whitehouse (now CTO, NCSC/GCHQ) on threat analysis. Now Senior Platform Engineer at Civica, where I champion AI adoption across the company, founded the general-outdoors channel, and recently led the team's first hike up Kinder Scout in Edale. Looking for the next thing worth building - Staff or Principal AI infrastructure, where the platform is the product.
An AI infrastructure engineer who builds tooling, scaling, and automation for modern engineering teams. Previously: PortSwigger, Civica. Now building at Stitch Studio.
Built, shipped, running
OpenClaw - the AI platform I run my life on
Designed and built a production AI operations platform: multi-model LLM gateway with failover routing, agent orchestration engine for parallel task delegation, 87 skill-based tool integrations, and semantic memory with RAG retrieval. Runs on a single VPS - because the right tool is the one that ships. Local voice pipeline - faster-whisper for speech-to-text, Piper TTS for text-to-speech. No cloud, no ElevenLabs, no third-party audio. Runs Etsy, YouTube, Notion, TikTok workflows - 24/7, zero human babysitting.
LLM Gateway - AI infrastructure backbone
Designed and built the central AI platform for a flagship security product. Every LLM call - chat, embeddings, RAG - routed through a single gateway. Model failover, rate limiting, cost attribution, audit logging. Multi-tenant, serving both product and internal tools. Stack: LiteLLM, Kubernetes (EKS), Helm, Prometheus, Grafana Loki.
Internal RAG system - before RAG was a category
Built a Q&A system over engineering documentation - vector embeddings, semantic search, retrieval-augmented generation - in early 2023, months before "RAG" entered the mainstream. Built for engineering; tech support found it and loved it. Nearly went live as the website's doc search.
Scan template generation - shipped in the product
Built template generation for vulnerability scans - describe what a scan needs to do, get a working template. Started as AWS Lambda PoC, moved to Kubernetes, shipped as an AI feature in the product. Users build vulnerability scans without hand-writing templates.
What I build when nobody's looking
micethateatcake - autonomous Etsy shop
A new digital product every day - designed, built, tested, and released by an AI pipeline driven by daily market research. Service contracts, onboarding kits, planner packs, printables - each day a fresh product researched against what's selling, built to match, and shipped to the shop. No human in the loop from research to publish. A real business operating at the edge of what autonomous systems can do today.
micethateatcake.etsy.com →Spotify Wrapped - year-in-review clone
A pixel-perfect rebuild of the Spotify Wrapped experience - animated data visualisation, shareable stat cards, the full slide-deck format. Built with Google Stitch Studio - exploring the space where design tools and production code meet.
wrapped.sushiroll.io →Job Hunting Board - networked job search engine
A structured job search dashboard that ingests LinkedIn CSV exports, maps your extended professional network, identifies warm introduction paths to target companies, and surfaces the highest-yield strategies. Tracks roles, companies, recruiters, scoring criteria, and intro chains - turning a senior-level search from chaos into a data-driven operation. Built with Playwright-based job board automation, deployed on Vercel with a client-side auth gate. Private, password-gated, updated daily with live roles.
demo preview → live board (pw) →Mora & the Fabulous Wonderfuls - band website + admin backend
Dark stage-lit portfolio for a Manchester live band. Full-bleed gig photography, masonry gallery, dynamic gig listings pulled from a custom admin API. Backend with drag-and-drop media uploads, WebP auto-conversion, EXIF stripping, gig CRUD, and storage quotas - all protected behind Caddy basic auth. Brand palette: Stage Black (#0B0E14), Stage Amber (#D4A030), Spot White (#E8E4DD). Barlow Condensed + Source Serif 4 typography.
mora.sushiroll.io →Nomic Atlas - AI tooling observability map
An explorable semantic map of 30+ tools, models, and infrastructure components powering the studio - embedded in Nomic Atlas with topic modeling and embedding-space visualisation. Covers LLM providers, content pipelines, deployment infrastructure, design tooling, and model registries. Built to demonstrate comfort with AI-native data tooling: PyArrow datasets, Nomic Embed for semantic indexing, and Atlas for interactive exploration. Shows a working mental model of a modern AI engineering stack.
demo overview →Howl's Bedroom - interactive art piece
An immersive recreation of Howl's bedroom from the Studio Ghibli film. CSS-drawn artefacts, ambient animation, no frameworks - pure HTML/CSS. Built as a study in detail and atmosphere. The kind of thing you build when you love the craft.
howlsbedroom.sushiroll.io →Tidbit Tails - AI young learners YouTube channel
A fully automated daily educational content operation on YouTube for young learners. Sora-2 animated Shorts, scripted, voiced, assembled, and published by an AI pipeline running on a content calendar. Each day a new video: researched topic → script → narration → generated clips → ffmpeg compositing → YouTube Shorts. Zero-touch from calendar entry to live. A study in autonomous daily content production at the edge of what generative video can do today.
@TidbitTails →Tidbit Tails - AI young learners TikTok channel
The TikTok arm of Tidbit Tails - same AI pipeline, formatted for TikTok's vertical Shorts feed. Each video is cropped and paced for the platform, published on the same content calendar schedule as YouTube. Automated from topic research through to live post, zero human touch per video.
@tidbittails →Mood board - Pinterest-style visual curator
A visual mood board builder - search by theme or keyword, get AI-curated image selections, lay them out in a Pinterest-style masonry grid. Built for product vision boards, brand aesthetics, and Etsy listing styling. A tool for thinking visually before building.
moodboard.sushiroll.io →Elysian - classical Greek showcase
An immersive Greco-Roman mythology experience - painterly temple scenes, vintage engraving illustrations, interactive hotspots revealing god myths. Built with AI-generated artwork and a design-system-first approach. A study in atmospheric web design.
classics.sushiroll.io →Get in touch
Currently Senior Platform Engineer at Civica, operating at Staff level. Open to Staff or Principal roles in AI Infrastructure / Platform Engineering. Remote-first (Manchester), flexible for London hybrid. If the work is interesting and the culture is genuine, I want to hear about it.
katie@sushiroll.io → grab my CV ↓