Euan Cowie

• Brought in on a fixed-term contract to de-risk an unreleased frontend/backend canvas product, architecting the execution layer that mediates between four moving parts: user intent, runtime preview, backend state and source code.
• Built a custom AST-based virtual file system modelling projects as round-trippable structured data, so imports, dependency-aware edits, code generation and refactors land in the right files instead of being patched as text.
• Built the iframe bridge between editor and runtime, Convex-backed projections for the inspector, and CRDT collaboration primitives; wrote custom agents operating over the project graph.
• Built an app-in-app shell in React Native and Expo — an outer application hosting and driving an inner one over a custom native bridge written for the purpose — with hand-written gesture and animation logic in Reanimated.

A scoped contract: brought in for a defined hard problem and finished it. The engagement ended when Bloom changed direction as a company.

• Architected a zero-touch hybrid-cloud control plane on AWS SSM and iPXE, cutting bare-metal provisioning from a week to under four hours and delivering software reliably to offline-first assets over Starlink and 5G.
• Separated deployment configuration out of the product into a versioned schema published by the monorepo and validated at the boundary, so a deployment that would not work is rejected before it is attempted rather than discovered on a vessel.
• Architected and led Infrastructure-as-Code across the company — Terraform with HCP, alongside CDK and CDKTF — over a hybrid estate spanning AWS and a colocation data centre, so engineers could self-serve while the company kept one view of the whole footprint.
• Built the network backbone reaching 15+ remote sites over cellular and satellite, with FortiGate HA paired to Transit Gateway VPN attachments and CloudWatch alerting, and led the research proving remote ingestion and inference viable over those links.
• Re-architected the real-time video pipeline — WebRTC and SRTP through an SFU, with NVIDIA hardware encode and decode — to restream from a single ingest instead of every consumer pulling its own, cutting network load and packet loss on bandwidth-starved satellite and cellular links.
• Ran GPU inference at the edge on NVIDIA Jetson devices, and built the multi-architecture build pipelines (Docker buildx on Jetson as native ARM capacity) that opened the Small Form Factor product line.
• Standardised GPU development environments for 20+ engineers on hosted VMs in the colocation estate, cutting build times from over 40 minutes to under 10.
• Directed DevOps strategy: deployment frequency up 300%, onboarding from days to minutes. Introduced test parallelisation and nightly soak-test safety checks — the product is life-safety-critical, so a regression reaching the field is not a support ticket.
• Technical lead for ISO drop-test certification, passed first time, which few do. Specified the deployment end to end — cabling, disk arrays, iSCSI, NVIDIA container runtime, firewall — and wrote the installation guides third parties followed onto cruise ships and oil rigs.

Cited inPulsant, “Revolutionising maritime safety with smarter systems”, April 2025 ↗

• Helped engineer and scale the "Golem Engine", a block-streaming GWAS kernel in CUDA that processes 500,000+ sample cohorts on a consumer NVIDIA 1080 Ti — optimising memory access patterns to stay inside an 11 GB VRAM ceiling while outperforming cloud equivalents.
• Designed and built the platform end to end: React client portals for researchers over a Rust/C++/gRPC control plane handling secure job dispatch and real-time result visualisation.
• Architected a two-container data-sovereignty model with a secure IPC protocol over Unix sockets, streaming encrypted results without exposing raw genomic data to the network.
• Built a fully self-hosted observability stack — DCGM exporter, Prometheus, Grafana, Loki — because genomic telemetry could not go to a hosted platform. Made VRAM high-water marks and workload imbalance across eight GPUs visible during runs rather than inferred afterwards: the same scheduling and memory-pressure problem as sharing a GPU host between training jobs.
• Spearheaded architecture across the platform, and mentored both the junior engineers and the academic staff who wrote code without having been trained as engineers.
• Migrated the system from a monolith to a service-oriented architecture, enabling the commercial launch of near-instant genome-wide association studies.

• Progressed to leading a cross-functional engineering team delivering the French tax law compliance module, handling high-volume transaction processing for Tier-1 private banking clients.
• Implemented the transition to a Spring and Kubernetes microservice architecture, decoupling critical reporting components to improve scalability and fault tolerance.
• Designed and implemented new API services for EU tax localisation, reducing integration time for client banking systems.

• Founded HEILSA LTD and built the whole thing alone: the iOS app in React Native and Expo, the backend, the knowledge graph, the infrastructure and the marketing site.
• Live on TestFlight with a small closed beta. My own five-person, multi-diet household has run entirely on it — pantry, meal planning, shopping — daily since February 2026, with eight external beta households onboarded.
• Built the Food Graph on Neo4j behind a Spring Boot service: graph RAG in the literal sense, with a vector index over ingredients and recipes for candidate retrieval and graph traversal over substitution, allergen and food-group edges for the answer. Ingredient identity resolves across pantry, recipes and shopping, so a tin of chopped tomatoes, a recipe calling for passata and a supermarket SKU reconcile to the same thing.
• Made recommendations fail closed: a recipe the graph cannot prove safe against a household of allergies and diets is not shown, every exclusion can be explained, and an answer is stored only if the preference version it was computed against is still current.
• Built multimodal extraction that turns a photo, a receipt or a typed sentence into structured pantry rows. Everything lands as candidates for review before it touches shared household state — an AI writing silently into a shared kitchen destroys trust the first time it is wrong — which also makes accept, edit and reject a free evaluation signal for extraction quality.
• Runs on a self-hosted Kubernetes cluster under Argo CD, with Prometheus, Grafana, Loki and Alertmanager in-cluster, plus Sentry and PostHog on the app side. Deliberately cheap to run, which is what keeps it a side project.

Deliberately small — self-funded, no employees, no outside money. It is how I keep my hands on the whole stack.

Live atheilsa.io ↗