Production RAG pipelines, edge-native infrastructure, and ML platforms at scale. I chose the engineering management path to drive change at the org level — but I drive the most value when I'm deep in the architecture and the code.
Form abandonment recovery for the rest of us. Captures partial submissions in real-time and sends magic-link recovery emails that restore the user's exact form state. 100% Cloudflare-native.
7 LoRA adapters across 4 model families for form abandonment classification. The best fine-tune hit F1 = 0.96 on a 6-class task where every general-purpose baseline scored 0.06. Dual-deployed on Modal for power and Cloudflare Workers AI for production inference.
A real-time pipeline debugger for the terminal, built in Rust. Drop it into any shell pipeline to see throughput, record samples, and format detection — without touching your data.
Infrastructure-as-code for my entire Cloudflare account. 8 domains, DNS, Workers KV, D1, Queues, R2, Vectorize, and AI Gateway — all managed via Terraform with zero secrets on disk.
Interactive visualiser for 25+ sorting algorithms with animated bars, step-by-step pseudocode highlighting, and sound mode. Built because time complexity isn't theoretical when you're processing 100k pages.
Find your next public holiday and optimise your leave for maximum time off. Supports 100+ countries with automatic location detection and a leave optimizer that maximises consecutive days off.
A single deep-research workflow spawned 107 sub-agents. Every one inherited Opus 4.8, because the fan-out passed no per-agent model. That one run cost about $89. Here's the PreToolUse hook I built so it can't happen silently again.
A rogue delete put a 200-resource CloudFormation stack into DELETE_FAILED. We took the surgical path instead of backup-and-restore and were back in two hours. Here's the technique and why AWS spreads it across two separate documents.
We were three months into a server-backed vector store redesign when AWS shipped S3 Vectors at re:Invent. Production traffic moved over the same day. Here's the story, the five options we evaluated, and what we'd pick again.
I fine-tuned 7 LoRA adapters across 4 model families for form abandonment classification. Every baseline scored 0.06. The best fine-tune hit 0.96. The interesting part was everything after.
AI tools are changing the skill distribution on engineering teams in ways most managers aren't ready for. The gap between "can produce code" and "can solve problems" is about to get very visible.