NSA-mini
I rebuilt DeepSeek's award-winning attention idea from scratch - the trick that lets AI models read very long text without slowing down. My GPU code runs the key part 22x faster than the standard approach.
mini-vLLM
The engine that runs AI models, built from scratch. It manages GPU memory in small pages so many people can chat with a model at once - nothing wasted, nobody waiting in line. Same core idea as vLLM.
Reasoning SLM
I trained my own small language model from zero - gathered the data, cleaned it, built the tokenizer, and taught a 118M-parameter model to reason about math and code. Every stage of the pipeline is mine.
An autonomous AI city. Agents live their own lives - they work, make friends, commit crimes, and stand trial. Nobody scripts them; the drama just emerges on its own.
Built solo, end-to-end. A Chrome extension that notices job listings while you browse and saves them to your dashboard automatically - no copy-pasting, no spreadsheets.
AttentionFM
A 24/7 AI radio station. AI hosts hold live conversations on any topic, around the clock - you tune in like a podcast that never stops.
Highway RL Agent
An AI driver that taught itself to drive through pure trial and error - crash, learn, repeat. Now it cruises highway traffic at ~80 km/h with zero collisions.
FinDocAgent
An AI that reads dense financial filings so you don't have to. Ask it a question, it answers with citations - so you can check every claim against the original document.
A service that watches stock prices in real time - it fetches live quotes, remembers them for instant answers, and streams every price change as it happens.
Send big files between your Mac and Android in seconds. Scan a QR code, drop the file, done - no cloud, no accounts, no ads.
Systemsbuilder,freshM.S.inhand.ItookanAIcampusassistantfromideato7,300+coursesserved,publishedresearchalongtheway-andnowIbuildtheenginesunderneath,fromscratch.
Research Software Engineer
University of Kansas
Built BabyJay, a production AI campus assistant serving 7,300+ courses and 2,207 faculty with 82.4% user approval - a multi-stage RAG pipeline (8 specialized retrievers, 35x faster retrieval), a FastAPI backend with JWT auth and 3-tier rate limiting, and 9 production data pipelines with Pytest CI regression checks. Mentored 100+ students as a graduate teaching assistant.
Software Engineer Intern
Note · USA (Remote)
Built Note, a developer-intelligence platform capturing and analyzing Claude Code sessions: 25+ REST endpoints in Next.js 16 App Router, a 15-table PostgreSQL schema with pg_trgm fuzzy search, a WebSocket + Redis pub/sub server for real-time CLI-to-web pairing, JWT dual-token auth, a 24-command Node.js CLI, and a 14-view React 19 dashboard.
Research Assistant
Amrita Vishwa Vidyapeetham
Co-authored 2 peer-reviewed papers (Springer LNNS ICT4SD 2024, IEEE i-PACT 2023) and shipped 2 production apps in Python, React, and PostgreSQL serving 500+ users; built a Bidirectional LSTM (491K parameters, TensorFlow) for next-word prediction and nutrient analysis.
DishKitPersonalized diet planning through next-word prediction and nutrient analysis
Long-term productivity and personal health depend on having a healthy diet. We propose an integrated dish preparation and ingredient delivery system using LSTM for predictive modeling, ML for nutrient calculation, and sentiment analysis to promote healthier eating.
SweepSpotMobile-enabled waste management platform with geolocation tracking
A waste management platform serving 500+ users across 10 municipalities with geolocation tracking, Haversine-based route optimization achieving 90% accuracy, and reporting dashboards for municipal authorities.
ConferencesPresented research on sustainable food delivery and waste management at international conferences
Let's build
something.
Got a project, a role, or just want to geek out about AI? I'm all ears.
~ ask me anything, or try one of these: