Computer Science · UIUC · Urbana-Champaign
I build reliable software systems and study how machine learning behaves in the real world—from LLM infrastructure to model evaluation.
Degree
Computer Science
University of Illinois
Graduation
May 2028
Academic record
3.83 / 4.00
Focus
Systems · ML research
Software Engineering Intern
Solo-built Alpha Engine: deterministic multi-asset signal research across crypto, US & Indian equities, F&O and forex — then backtested it honestly and published the null result rather than a claim.
Built a deterministic research engine — 22 independent analyzers synthesized into directional signals with calibrated confidence and explicit invalidation prices (33K LOC, solo).
May 2026 – Aug 2026
Software Engineer (part-time) — HDF5 AI Pipeline
LLM refactoring engine for the 25-year-old HDF5 C library with a dual-LLM consensus gate (Claude + GPT-4o both_approve/disagree).
Built the primary LLM refactoring engine (457 LOC) generating git-format-patch-validated diffs with invariants enforcing no-new-API / no-new-export / no-new-file.
May 2026 – Aug 2026
Machine Learning Engineer
Owned the data layer of a DistilBERT classifier that detects FDA/FTC advertising compliance violations across 15 regulatory categories.
Built the preprocessing and tokenization infrastructure turning raw advertising scripts into model-ready tensor datasets (input_ids, attention_mask, labels) in PyTorch and HuggingFace Transformers.
Jan 2025 – May 2026
Junior AI & NLP Lead Researcher
Led a 31-intern team (Agile) improving a production exam platform.
Led a cross-functional team of 31 interns with Agile methodology, improving exam model accuracy by 15%.
May 2024 – Aug 2024
Machine Learning Intern
Shipped three production AI systems end to end.
Jun 2023 – Aug 2023
AI/ML Program — Top 20 Nationwide
Selected among 20 students nationwide for an intensive AI/ML program at IIT Delhi's IHFC innovation hub.
2022 – 2023
Leadership & service
Selected, in addition to engineering roles
Founder & President · 200+ volunteers · 6 cities · $24K
Founder & President · 80+ members · $7K raised
Secretary General ×2 · 35+ MUNs · 17 Best Delegate
An autonomous LLM cost-governance proxy that sits in front of OpenAI and Anthropic. It enforces per-project and per-feature spend ceilings on the hot path — authenticate → attribute → predict → circuit-break → reserve → forward → capture — while a caller's existing provider SDK keeps working unmodified.
FastAPI metering proxy with an authorize/capture reservation system under a single asyncio lock, dual-protocol SSE usage parsing, and a two-condition circuit breaker.
A local MCP server that gives AI coding assistants graph-aware memory — it retrieves whole dependency chains (callers and callees), not isolated files, so the model sees the real context around a symbol.
Local MCP daemon: hybrid FAISS HNSW + SQLite FTS5/BM25 fused with RRF (k=60), plus 2-hop BFS call-graph expansion.
A real-time, bidirectional ASL ↔ speech translator: sign-to-speech and speech-to-signed-3D-avatars, running dual-hand tracking in the browser at ~30 FPS.
Bidirectional ASL ↔ speech: MediaPipe dual-hand tracking → TF.js classifier → Whisper/Gemini/ElevenLabs pipeline, 3D avatars.
A full-stack adaptive learning platform whose multi-agent backend personalizes every step — powered by a Bayesian Knowledge Tracing engine written from scratch.
Multi-agent backend (7 ReAct agents) + a Bayesian Knowledge Tracing engine built from scratch that picks questions by expected information gain.
More project detail, implementation notes, and the interactive F1 presentation are available in the F1 version.
Reproduced a published protein-structure-classification benchmark across 72 datasets in Python/pandas/Matplotlib — finding traditional ML stays competitive with deep learning at a fraction of the runtime — plus a counterfactual name-perturbation replication of an NLP toxicity-bias workflow.
Compared ChatGPT, Gemini, Perplexity, and Claude using structured factual, open-ended, situational, hypothetical, and creative prompts, with quantitative and qualitative review. The paper reports Gemini higher on factual accuracy and ChatGPT stronger on creative text generation in its examples.
Best Junior Author of the Year ($1,000) · most-downloaded high-school-authored paper in journal history.
The 2023 paper and patent need a verified public record or identifier before full citation details can be added.
Technical areas
From project and course work
Languages · Python · TypeScript · Rust · Kotlin · JavaScript · C++ · SQL · Java
AI & ML · Multi-Agent Systems · Transformers · PyTorch · BKT · MediaPipe · TensorFlow.js · HuggingFace · LLM Pipelines
Full-Stack · Next.js · React · FastAPI · Node.js · Supabase · Docker · Tailwind · Three.js
Data & Infra · pgvector · FAISS · SQLite · ETL Pipelines · Cloud Run · GitHub Actions · OAuth2
Research · Explainable AI · NLP · Bayesian Methods · Technical Writing · Experiment Design