I build production AI agent systems and the evaluation harnesses that catch them when they quietly stop working.
Hello! I'm Shreeja, an MS Computer Science candidate at Illinois Institute of Technology (GPA 3.6/4.0, graduating May 2027) with 1+ year of industry experience building AI systems that run in production. Currently an AI Engineer Co-op at Aion Labs, where I've shipped 16+ peer-reviewed production PRs across a 3-repo AI data platform — fixing production defects in a Three.js/React 3D engine, building data-quality tools that now gate pipeline releases, and running the platform's first live data refreshes in 75+ days. I'm also a Claude Certified Developer (Anthropic).
Previously a Software Engineer (Quality) at Microchip Technology, where I built an NLP-powered chatbot integrated across ASPICE, JIRA, POLARION and CAD that cut manual search time by 40% for a 50-engineer team.
What I care most about is whether systems actually work. My data-quality checker caught a million-fold unit error in production data on its very first run; on a multi-agent project, I found my own LLM judge was scoring leniently and masking real regressions. Catching what everyone else assumed was fine is the work I find most satisfying.
An 8-agent routing graph with LLM intent classification directing queries across 4 specialist paths (research, teaching, coding, comparison), grounded by a ChromaDB vector store and real-time web search, deployed through GitHub Actions CI/CD. I built an LLM-as-judge evaluation framework over 5 benchmarked query types reaching 8.0/10 quality at 4.72s latency — then found the judge was scoring leniently and hiding real regressions. Diagnosing and fixing that bias was the most valuable thing I did on the project.
↗A stock research assistant built on the Model Context Protocol. Rather than one monolithic script, it's a three-server architecture — market data, research, analysis — with a single host holding a dedicated client per server, following the protocol's one-to-one client/server model. The market-data server ships with typed tool schemas auto-derived from Pydantic field definitions via FastMCP, validated in MCP Inspector. Retrieval and evaluation layers are in active development.
↗With a five-person team, a 7-phase distributed PySpark pipeline over 6.96M records: cleaning, stratified splits, 8 engineered features, Parquet persistence, profiled with Athena on AWS EMR. I owned the TensorFlow DNN track — and gradient boosting won, at 84.67% accuracy and 79.14% F1 against the DNN's 81.45%, training 3.3× faster. Worth knowing when the simpler model is the right call.
Reach out through any of the platforms below. I'm graduating in May 2027 and open to AI/ML and software engineering roles.