Applied ML Engineer · MS Applied Data Science @ USC · Research Assistant @ USC CeNEC Lab
I build ML systems end to end — from quantizing a model onto an NPU, to training a policy with RL, to auditing whether a paper's code actually reproduces its claims — and I write down what failed along the way.
Open to: Applied ML Engineer · Research Engineer · ML / AI Engineer · Data Scientist (ML/NLP/CV)
| Project | What it is | Stack |
|---|---|---|
| Snapdragon Edge Access Control | Quantized ArcFace face recognition compiled for the Snapdragon X Elite NPU — 1.71 ms/face, 188/188 layers on NPU — driving an Arduino Uno Q access indicator. No camera frame leaves the device. | PyTorch · ONNX · QNN · Qualcomm AI Hub · Arduino |
| RepoRaven | Agentic auditor that checks ML papers against their own code and cites the exact line that contradicts the paper. Found real data leakage in G-Mixup (ICML 2022). | Python · LLM agents |
| School Facilities from Imagery | Measures school facilities from overhead imagery with a calibrated confidence on every field; open-weight models on SLURM, leave-one-school-out calibration, offline reproducible. | PyTorch · open-weight VLMs · SLURM |
| nika_ai | From-scratch decoder-only transformer with a routed MoE MLP, hand-rolled ring all-reduce vs. real DDP, and a torch.export investigation of MoE graph breaks. 18 tests. |
PyTorch · torch.distributed |
| Tau-bench RLVR | GRPO fine-tuning of Qwen2.5-7B (4-bit + LoRA) on tau2-bench with rule-based verifiable rewards for structured tool calls. | TRL · PEFT · bitsandbytes |
| RL Autonomous Driving | PPO agent that learns to drive a custom ray-cast simulator; documents every reward-shaping iteration, including the ones that failed. | Stable-Baselines3 · PyGame |
| Intelligent Bistro | Conversational food-ordering app — voice, text, or tap — with native Claude tool calling, SSE streaming, and 159 tests in CI. | TypeScript · React Native · Node |
| ImmunyAI | Allergy tracker combining MedGemma (VLM), simulated wearable vitals monitoring, and a voice UI. | React · TypeScript · MedGemma |
More: Insurance policy Q&A on Cloudflare Workers · Immune repertoire profiling from RNA-seq · Spatial staging: a Blender you talk to
Research Assistant — CeNEC Lab, University of Southern California · Present
- LLM benchmarking research comparing multimodal models (Gemini 1.5/2.0/2.5, GPT-4/5) against human performance on reasoning and perception tasks
- Designing evaluation datasets and protocols; RAG-based alignment for cognitive comparability
- Statistical validation across 4+ state-of-the-art models
Machine Learning Engineer — Mirador Technologies · 2025
- S.I.C.A.I.R. — SAR image colorization framework using SAM and CycleGAN for context-aware, computationally inexpensive colorization
- Re-architected a production CV pipeline, improving face recognition accuracy by 51%
- Shipped 7+ computer vision and NLP projects with ≥87% accuracy; optimized inference to <100 ms latency
- Diabetic Retinopathy Detection from Retinal Images — CNN-based early detection from medical imaging
- Contextual Sentiment Analysis Using BERT-based Models — fine-tuning for domain-specific sentiment classification
- Sentiment-Driven Portfolio Optimizer — FinBERT sentiment scores feeding Black-Litterman expected returns
All publications and achievements →
- ML / DL: PyTorch, TensorFlow, JAX, Hugging Face Transformers / TRL / PEFT, Stable-Baselines3, scikit-learn, XGBoost, LightGBM, OpenCV, FAISS
- LLMs: fine-tuning (LoRA, GRPO/RLVR), RAG, tool calling, agents (LangChain, LangGraph), evaluation and benchmarking; Llama, Qwen, Gemini, GPT, Claude, MedGemma
- Edge & systems: ONNX, Qualcomm AI Hub / QNN, quantization, CUDA,
torch.distributed, SLURM, Docker - Languages: Python, TypeScript, C++, SQL, R, MATLAB
- Data & cloud: Pandas, NumPy, Spark, PostgreSQL, MongoDB, Snowflake, Firebase, AWS, GCP, Cloudflare Workers, FastAPI