MS Data Science · UC San Diego  |  GPA 4.0
Applied ML  ·  Inference Optimization  ·  On-Device AI

LinkedIn  ·  HuggingFace  ·  kdcosta@ucsd.edu

I build ML systems end-to-end — from research and experimentation to deployment on real hardware. My work spans model evaluation, inference optimization, data pipelines, and on-device AI across mobile, x86, and GPU targets.

Previously: 1.5 years at IIT Bombay's Vision & Image Processing Lab. Three IEEE publications. One published patent. Currently graduate researcher at MOSAIC Lab, UCSD.

Reproducible benchmarking framework for 7 GGUF variants of Llama 3.2 3B and Qwen 2.5 1.5B across Pixel 6a (ARM), Intel x86, and Apple M4 (Metal GPU).

• Thermal isolation protocol reduced trial variance from ±8% → ±2%
• Q2_K runs 112% faster than Q6_K on ARM; Metal GPU reverses the ranking entirely
• Modeled KV-cache throughput cliff; predicted x86 collapse within 8%
• Published dataset on HuggingFace · Live interactive dashboard with CI/CD

→ Dashboard  ·  → Dataset

Python llama.cpp ADB CI/CD

Production-grade recommender system on 1.1M Food.com interactions. Time-aware matrix factorization with bias terms to model 18-year rating drift. Deployed as a REST API with semantic search and full MLOps stack.

• RMSE 0.68  ·  NDCG@10 0.23
• FastAPI + FAISS vector search · Deployed on Google Cloud Run
• 162 tests · CI/CD pipeline · Docker containerized

Python FastAPI FAISS scikit-learn Docker GCP

ML competition — predict coastal flooding across 12 tide gauge stations with OOD generalization as the core challenge. Trained on 9 stations, evaluated on 3 unseen. Modular src/ pipeline with SQL feature engineering, fixed seeds, MCC primary metric.

Python XGBoost SQL scikit-learn

Curated and published a unified multimodal retinal dataset (~4,700 images, 8 sources) on Kaggle. Two-stage ensemble: custom CNN + transfer learning (ResNet50, VGG16, EfficientNetB0, DenseNet121)

• logistic regression meta-learner.

• OCT: 99.41% accuracy  ·  Fundus: 97.70% accuracy
• Dataset DOI: 10.34740/kaggle/ds/7394927
• Accepted: IEEE ICCCNT 2025

Python TensorFlow OpenCV Kaggle

Hackathon project built on AWS and Databricks. Contributed SnapTrash — a geospatial feature for crowdsourced waste mapping with location-tagged image classification.

AWS Databricks Python

ML/AI: PyTorch · TensorFlow · scikit-learn · ONNX · llama.cpp · Transformers
Languages: Python · Java · SQL · JavaScript
Systems: Android (Java/Kotlin) · Flask · Docker · ADB
Cloud: AWS · GCP · Databricks
Tools: Git · HuggingFace · Kaggle · OpenCV