Backend engineer
I work on the part of software nobody sees until it breaks: the APIs, services, and data underneath everything else. Lately, at work and after hours, I've been bringing LLMs down there too, where answers have to be right, not just fluent.
A fact-checking agent written in Python from scratch — no LangChain, no LangGraph. Give it a claim and it searches the web through Tavily, then returns a structured JSON verdict with citations. Built to understand the raw agent loop before reaching for abstractions.
Python · Gemini · Tavily
GitHubA modular research pipeline that chains three LLM-driven agents to turn a raw topic into a structured, decision-ready report. Built for analysts who want usable insight without writing prompts or API boilerplate.
Python · Ollama · Cerebras
GitHubA developer toolkit that keeps a dozen small utilities in one tab, so you stop hunting for the site that formats JSON or decodes a JWT. Private repository.
React · JavaScript · Tailwind CSS
Live siteAn LDAP authentication bypass lets anyone log in when the bind is not done with the user's own credentials. Why empty-password logins work, and how to fix it.
QUERY is now a published HTTP standard. Here is what it actually does, what happened when I sent real QUERY requests from curl and Java 21, what the Spring source says about handling one, and whether you should use it yet.
Build the second phase of Trinetra: an AI agent that extracts every claim from a paragraph, fact-checks each one, and renders a Claims Atlas. Plus the model-choice lesson that turned a wrong verdict into a right one.
Building retrieval-augmented generation in Spring AI with Gemini and an in-memory vector store. The four dependencies nobody warns you about, the embedding auth gotcha, and a real example of similarity search ranking the wrong document first.
How portable is Spring AI really? I ran the same triager code against Gemini and OpenRouter to find out where the abstraction holds and where it leaks. Maven profiles, the ambiguous-bean trap, and the difference between two kinds of portability.
I've spent about a decade across the stack — designing REST and GraphQL APIs, shaping database schemas, and building the React and Next.js front ends that sit on top of them. Most of my recent work has been on LLM systems: agent loops, retrieval, and tool calling, usually in Python or Spring AI.
When I'm not writing code I'm writing about it. The blog covers what I learned the hard way — system design, prompting, and the tools I keep coming back to. I'm open to full-time roles and select freelance work.