Turns network change validation — checking nothing broke after a config update — from a multi-week manual process into a 10-minute automated check, catching silent failures a human reviewer would miss.
Reads vendor device documentation and automatically writes and runs the test suite needed to certify new network hardware, replacing a weeks-long contractor process with one that finishes in minutes.
A team of specialized AI agents — one supervisor, several specialists — that monitor and troubleshoot infrastructure together, cutting the manual troubleshooting operators do by 40%.
A live map of how devices and systems depend on each other, so operations teams can see the blast radius of a planned change or an outage before it happens, not after.
A data pipeline that pulls operational metrics from many tools into one governed data lake with live dashboards, cutting the time stakeholders wait for reliable data by over a third.
A chat-based operations assistant that answers questions from internal knowledge instead of generic AI answers, and scales each part independently as usage grows.
A flight recorder for AI coding agents working in an AWS account: it shows every change an agent made, flags the risky ones, lists what it left running and what that costs, and drafts the IAM guardrails to stop it next time.
Read the write-up →A durable AI on-call agent that investigates incidents in a simulated production environment, built to test, with real measurements, whether small helper models should handle an agent's context housekeeping.
Read the write-up →A personal AI workspace where you plug in your own API key to chat, validate business ideas, and connect real tools (GitHub, Grafana, files) — instead of paying for a locked-down AI subscription.
An AI investigator for outages — feed it logs, traces, and metrics and it builds a cited, evidence-backed timeline of what went wrong, instead of a wall of raw data an engineer has to piece together by hand.
A from-scratch reliability monitoring system that tracks whether a service is meeting its uptime promises and automatically pages the team when it's burning through its error budget too fast.
Read the write-up →A research benchmark proving an AI search system can run entirely on a local machine, with no cloud API calls, while cutting the amount of text it needs to read by nearly 1,000x with no loss in accuracy.
A controlled experiment testing which AI-agent design choices — planning, memory, verification — actually make agents more reliable, so engineering effort goes toward what measurably works instead of guesswork.
A serverless system that screens card transactions for fraud in real time, modeled on the kind of low-latency infrastructure banks and fintech companies rely on to catch bad transactions in milliseconds.
A hands-on run through AWS's end-to-end SageMaker AI workshop, followed by my own experiments showing how data leakage and the wrong train/test split make a model look far better than it is.
Read the write-up →Hands-on labs fine-tuning open models four different ways, and a written review of which technique fits which problem, so customization choices are made on evidence instead of trial and error.