Sophomore studying Artificial Intelligence at UC San Diego (GPA 3.98), working across machine learning research, embedded inference, computer vision, and full-stack AI systems.
Active LearningEMG / BiosignalsComputer VisionLLM Fine-TuningRAG SystemsEdge DeploymentPyTorchReact.js + Flask
About
The person behind the models
3.98
GPA at UC San Diego
6+
Research & project roles
63%
Relative improvement in autonomous CI/CD agent pass rate
7+
Certifications across ML, AI & GPU computing
I'm a second-year Artificial Intelligence student at UC San Diego, working at the intersection of ML research, applied systems, and real-world deployment. From bioacoustic classification pipelines to EMG-based gesture interfaces, I care about AI that actually runs in the world — not just in notebooks.
My current research spans machine learning for bioacoustic species identification — including a new RAG-conditioned framework that retrieves acoustically similar exemplars to improve rare-species classification — at Engineers 4 Exploration, and EMG-based ASL gesture recognition for neurotechnology at Triton NeuroTech. Outside of research, I'm building an autonomous CI/CD repair agent that detects failing tests, generates and applies patches, and reruns pipelines with zero human intervention. I also mentor students on AI-driven projects through Curious Cardinals.
Before college I participated in ML research at the Cambridge Centre for International Research alongside Dr. Shadi Ghiasi, deployed a Raspberry Pi fire-detection system with a UC Irvine CS professor, and completed ML coursework through Stanford and Coursera.
Always looking for research collaborations and internship opportunities in ML systems, applied AI, and edge deployment.
Experience
Where I've worked & researched
Jan 2026 — Present
Machine Learning Research — Acoustic Species ID
UC San Diego: Engineers 4 Exploration
Building a machine learning pipeline for multi-label bioacoustic classification using BirdNET transformer embeddings with LightGBM, Random Forest, SVM, and MLP classifiers. Achieved ~90% precision on dominant species classes. Applying uncertainty-based active learning sampling to maximize model efficiency under limited annotation budgets across imbalanced long-tail distributions. Designed and evaluated a RAG-conditioned framework for NatureLM that retrieves acoustically similar exemplars via BirdNET/BEATs embeddings to improve rare-species classification accuracy, with statistically validated gains over baseline (McNemar's test, bootstrap CIs).
Engineering EMG-based ASL gesture recognition models translating raw muscle biosignals into structured linguistic outputs. Extracted time-domain (RMS, MAV) and frequency-domain spectral features for robust multi-class classification. Achieved ~95% test accuracy with an optimized Random Forest classifier. Preparing pipeline for real-time edge deployment with low-latency inference.
EMG Signal ProcessingASL RecognitionNeurotechnologyEdge Deployment
Jun 2026 — Aug 2026
Autonomous CI/CD Repair Agent
Personal Project
Architected an end-to-end agentic pipeline where GitHub Actions test failures trigger a FastAPI webhook server that feeds failing logs to an LLM, generates a patch, applies it, reruns tests, and opens a fix PR automatically — a fully autonomous agent with zero human in the loop. Improved automated fix success rate by 63% on SWE-bench-style test cases, raising the agent's pass rate from 14% to 23%. Currently extending with structured repository-context retrieval (file-tree parsing, import-graph traversal) as a precursor to embedding-based semantic retrieval and multi-agent orchestration.
Trained a YOLO-based satellite imagery detection model achieving 92.5% accuracy on swimming pool identification for large-scale geospatial analysis. Improved model robustness by annotating ambiguous visual edge cases to reduce false positive rate in production inference.
YOLOComputer VisionSatellite ImageryGeospatial ML
Jan 2026 — Present
AI Mentor
Curious Cardinals · Remote
Mentoring students on AI-driven coding projects from concept to working ML pipeline. Teaching ML fundamentals, model design, and debugging. Helping students translate abstract ideas into functional systems through iteration and experimentation.
AI EducationMentoringML Systems
Dec 2025 — Jan 2026
Machine Learning Researcher
RapidFire AI · Remote
Fine-tuned GPT-2 with LoRA adapters via SFT for customer-support generation (ROUGE-L=0.056, eval loss=1.15). Built a separate RAG pipeline on the FiQA financial dataset, running a grid search over chunking strategy and reranking configuration to optimize retrieval quality, achieving MRR=65.5%. Designed parallelized hyperparameter search with real-time metric visualization and early stopping.
Conducted interview-driven user research and NLP-based sentiment analysis to identify product gaps. Recommended integration of a proactive AI forecasting feature. Benchmarked ML algorithms across multi-year financial time-series, competitor metrics, and news-driven datasets.
NLPSentiment AnalysisFinancial ML
May – Aug 2024
Embedded Systems Research
Pioneer Academics × UC Irvine CS Faculty
Collaborated with a UC Irvine CS professor to engineer a smart fire-detection camera system. Trained YOLO-based networks achieving 90.2% accuracy on fire and 85.7% on smoke. Deployed end-to-end on Raspberry Pi with on-device inference and IoT-based local alert transmission pipeline.
Participated in a formal ML research program alongside Dr. Shadi Ghiasi, a senior AI research specialist. Authored a research paper applying machine learning to ECG-based diabetes detection. Compared 16+ classical and ensemble models, improving Random Forest precision to 87.5% through feature engineering and signal processing. (Academic research program — not a peer-reviewed publication.)
Chrome extension combining real-time eye-tracking gaze signals with an event-driven LLM + TTS pipeline. Gaze-to-DOM coordinate mapping fires contextual explanations based on exactly what you're reading. Debounce logic and async optimization minimize unnecessary API calls.
Fine-tuned GPT-2 with LoRA adapters via SFT for customer-support generation. Separately, built a RAG pipeline on the FiQA financial QA dataset with optimized chunking, overlap, and reranking, evaluated via Precision, Recall, F1, NDCG@5, and MRR. Parallelized hyperparameter search with early stopping across both SFT and RAG experiments.
End-to-end agentic pipeline: GitHub Actions test failures trigger a FastAPI webhook, which feeds failing logs to an LLM that generates a patch, applies it, reruns tests, and opens a fix PR automatically — a fully autonomous agent with no human in the loop. Currently extending with structured repository-context retrieval as a precursor to multi-agent orchestration.
Full-stack platform aggregating Canvas LMS and Gmail into a unified academic dashboard. Built the LLM integration layer that parses raw API data, applies a deterministic pre-filter to block junk without unnecessary model calls, and surfaces actionable updates in real time. React.js frontend with Flask backend, OAuth auth, and Supabase for multi-service data ingestion.
YOLO-based neural network deployed end-to-end on a Raspberry Pi for real-time fire and smoke detection. Full IoT alert pipeline notifies local residents on detection. Developed in collaboration with a UC Irvine CS professor.
AI-powered ocean debris coordination platform built in 36 hours. Field sightings captured via ElevenLabs voice input and analyzed on-site by a pretrained Roboflow CV model to classify severity; Groq (LLaMA 3.1) handles triage, crew assignment, and response recommendations. Drift forecasting combines CORC Spray glider data, HYCOM fallback, and Scripps/NOAA oceanographic data to predict trajectories at 24/48/72-hour windows. Live ops dashboard with AI-ranked crew dispatch across ships and shore teams, synced via Supabase Realtime.
Cambridge Centre for International Research · 2023
Comparative ML study across 16+ models for diabetes detection using ECG-derived features, conducted through a research program with Dr. Shadi Ghiasi. Applied feature engineering and signal processing to benchmark and optimize classical and ensemble methods.
RAB for NatureLM: Retrieval-Augmented Bioacoustic Classification
Introduces a RAG layer for NatureLM's audio-LLM bioacoustic classifier, comparing BirdNET and BEATs-based retrieval across 770 clips of 77 North American bird species. Finds retrieval source matters more than retrieval mechanism, accuracy gains plateau near k≈7 candidates, and the model is roughly 5x more sensitive to retrieved text context than to audio.
Smart Fire & Smoke Detection via Embedded IoT Cameras
Describes a YOLO-based fire and smoke detection system deployed end-to-end on Raspberry Pi hardware, achieving 90.2% fire and 85.7% smoke detection accuracy with on-device inference and an IoT alert pipeline.
ECG-Based Diabetes Detection: A Comparative Model Study
Compares 16+ classical and ensemble machine learning models for diabetes detection using ECG-derived features, improving Random Forest precision to 87.5% through feature engineering and signal processing.
AI is what I do technically — but this is what shaped how I think, lead, and show up for people.
🏆
National Winner — FlexFactor
NextFlex · Jun 2023
Led team to a national victory out of 350+ participating teams. Presented a full technology product design, business model, and pitch deck to government officials, industry leaders, and academics. Recognized for innovation, teamwork, and entrepreneurial thinking.
🧠
Neuroscience Summer Program
UCLA · Summer 2024
Studied brain anatomy, neurological disease, and AI-assisted stroke detection with UCLA faculty mentorship. Directly shaped interest in neurotechnology — which led to EMG/ASL research at Triton NeuroTech.
🌲
Co-Founder — Adopt a Park
Feb 2022 – Jun 2025 · 3+ years
Co-founded and led a community environmental initiative organizing volunteer efforts and park maintenance at Glenview Park, San Jose over more than three years.
⚜️
Eagle Scout — Boy Scouts of America
Aug 2018 – Jun 2025 · ~7 years
Earned the rank of Eagle Scout, BSA's highest honor, achieved by fewer than 6% of Scouts. Planned and led a capstone community service project from proposal through execution, while mentoring younger scouts across multi-day service and leadership activities.
🤝
AI Mentor — Curious Cardinals
Jan 2026 – Present
Mentoring students on AI-driven projects from concept to working implementation. Teaching ML fundamentals while helping students debug, iterate, and think like engineers.
📜
Languages
English (native) · Sanskrit (working proficiency)
Native fluency in English. Working proficiency in Sanskrit — one of the world's oldest structured languages with deep connections to formal grammar and computational linguistics.
Contact
Let's work together
I'm always open to research collaborations, internship opportunities, and conversations about ML systems, applied AI, and edge deployment. Feel free to reach out through any of the channels below.
