Dev said yes. Production said prove it. That's where most of what I know comes from — systems meeting reality, quiet assumptions becoming loud problems, and the engineering that happens after the first design breaks. I write about the proof.
Read the articles →Software engineer at Vertex Inc. — building across the full stack on AWS, deploying with Kubernetes and Terraform, debugging production systems that worked perfectly in dev.
In grad school, a research assistantship in a lab partnered with the steel industry led to two published papers on computer vision for meltshop safety. An unexpected detour that taught me how software meets the physical world.
Building a RAG system without vectors just to see if it works. Wiring Playwright to Hacker News to build a personal AI journalist. Training agents to turn safety logs into compliance reports.
Writing about the production lessons that textbooks skip. Exploring knowledge graphs, ontologies, eBPF, and whatever else won't let me sleep until I understand it. Debugging with snacks.
Same instinct in both: build it, hit a wall, write down what it taught me.
Scaling doesn't break systems — it exposes the assumptions we didn't know we were making. Cache worked in dev, worked with one pod, then fell apart at scale.
Read more →If users experience your system as unpredictable, they'll work around it — and that's when security actually breaks down.
Read more →Retrieval architecture matters as much as the LLM itself. An experiment in graph-based retrieval — from regex to NLP to LLM-based entity extraction.
Read more →Everything worked in Chrome and Edge. Safari blocked the API calls completely. The root cause wasn't service logic — it was CORS gaps at the API gateway level.
Read more →Before a system can be intelligent, it first needs to understand the world it's operating in. Taxonomy tells us what things are. Ontology tells us how things relate.
Read more →Move intelligence closer to where decisions are made. Make complex systems programmable. Cloud, networking, and kernels — the same philosophy applied differently.
Read more →Constraints often point to better boundaries in system design. A hard payload limit led to a better architecture — async, observable, and resilient.
Read more →Machine-to-machine clients don't use ID tokens. API Gateway's native authorizer refuses to validate Access Tokens — a quiet gap that blocks integration flows.
Read more →Rate limiting isn't about saying "no" — it's about keeping your system stable enough to keep saying "yes." And in production, it's always layered.
Read more →Detection worked well, but huge volumes of logs are extremely hard to manually convert into meaningful insights. Multi-agent orchestration bridges that gap.
Read more →From topic to insight to audio. Browser automation + MCP tools + AI models working together to reduce information overload.
Read more →If you're not reading the plan, you're mostly guessing. Treating performance issues as system behavior made debugging more predictable.
Read more →You can't optimize history, performance, and duplication at the same time. Something has to give — and avoiding duplication often pushes complexity elsewhere.
Read more →Observability isn't about more dashboards — it's about faster answers. Deployments went from hopeful to verifiable.
Read more →3,280 links evaluated — 81% rejected. Same link, different page, different answer. Context-sensitivity was the most surprising behavior.
Read more →Graph-based retrieval without embeddings or vector stores. Three-phase entity extraction (regex, spaCy, LLM) building a knowledge graph from AWS documentation.
View on GitHubMulti-agent AI pipeline that transforms PPE safety logs into stakeholder-ready compliance reports. Data Preparer, Analyst, Reporter — each with prompt-driven behavior contracts.
View on GitHubTopic-to-audio pipeline: scrapes Google News and Hacker News, generates summaries with AI, converts to speech with ElevenLabs. Playwright + MCP + LLM integration.
View on GitHubCNN-based model that recognizes sign language alphabet gestures via webcam in real time. Built with Keras — an early exploration of computer vision for accessibility.
View on GitHubYOLO-based vehicle type classification with Haar cascade for speed analysis. Real-time detection from video feeds — bridging object detection and traffic analytics.
View on GitHubSymptom-based disease prediction using web-scraped medical data and a decision tree classifier. Practical ML applied to health — from raw web data to diagnostic suggestions.
View on GitHub