I build AI systems end-to-end from raw data ingestion to model inference to agentic workflows, and I care about what happens to the data before it reaches the model.

Most AI engineers treat infrastructure as someone else's problem. I don't. My background spans multimodal ML, real-time streaming pipelines (Azure EventHub → Databricks → Delta Lake), and LLM-powered agentic systems, all built on production-grade data foundations.

Current focus: Healthcare AI, agentic systems with RAG + tool calling, and LLM fine-tuning under resource constraints (PEFT/LoRA, 4-bit quantization).

PyTorch · Hugging Face · BiomedCLIP · Meditron-7B · Clinical-T5 · MedLLaMA · PEFT/LoRA · CUDA/DDP

Production-scale multimodal AI system on 119,533 MIMIC-CXR records. Built for two simultaneous clinical tasks: generate free-text radiology reports from chest X-rays, and classify 14-label pathology findings under severe class imbalance.

What makes this non-trivial:

• Designed a CLS-conditioned cross-attention abstractor that compresses 197 ViT tokens → 32 query tokens, enabling efficient vision-language fusion while training only 73M of 3.6B parameters
• Applied PEFT/LoRA + 4-bit NF4 quantization + CUDA/DDP — full-scale training on constrained hardware without sacrificing model capability
• Engineered clinical preprocessing with OpenCV CLAHE, schema validation, deduplication, label normalization, and patient-stratified splits to prevent data leakage across train/val/test
• Built threshold-tuning workflows with AUROC, F1, BLEU-4, METEOR, and ROUGE-L; achieved 0.830 micro-AUROC and improved rare-class sensitivity by 70.88%

Research output: Accepted abstract at SPIE Medical Imaging (peer-reviewed)

Gemini APIs · LangChain · RAG · Azure EventHub/Kafka · Databricks DLT · Spark/PySpark · Delta Lake · Unity Catalog · Streamlit

A natural-language analytics assistant built on top of a production real-time streaming pipeline. Ask questions about pricing, demand, city-level KPIs — the agent retrieves, reasons, and answers against live data.

Pipeline architecture:

• Real-time ingestion of ride events across 40 US cities via Azure EventHub (Kafka-compatible)
• Databricks Delta Live Tables with Medallion Architecture (Bronze → Silver → Gold), Star Schema, SCD Type 1/2, Unity Catalog
• Partition tuning cut DLT execution from 6–8 min → 75 seconds (85% reduction)
• Azure Key Vault for SAS key rotation; Python SDK for automated EventHub lifecycle management

Agentic layer:

• Gemini APIs + LangChain-style orchestration with RAG, tool calling, and MCP-style tool interfaces
• Structured JSON outputs, prompt validation, and self-healing pipeline checks for schema drift and failed transformations
• Streamlit dashboard with KPIs, Plotly visualizations, and Folium heatmap for city-level demand

AWS S3 · Snowflake · dbt Core · Streamlit

ELT pipeline with IAM-authenticated S3 → Snowflake external stages, dbt Core transformations (Jinja macros, SCD Type 2 snapshots, schema tests), and Medallion Architecture (Bronze → Silver → Gold). Live Streamlit dashboard covering pricing, occupancy, and host metrics.

AI / ML         PyTorch · Hugging Face · BiomedCLIP · Meditron-7B · Clinical-T5 · MedLLaMA
                PEFT/LoRA · 4-bit NF4 · CUDA/DDP · scikit-learn · XGBoost · OpenCV

Agentic / LLM   LangChain · RAG · Tool Calling · MCP-style Interfaces · Gemini/GPT APIs
                Structured JSON Outputs · Agent Routing · Self-Healing Workflows

Data Eng        Apache Spark/PySpark · Databricks DLT · Delta Lake · Unity Catalog
                dbt Core (Jinja macros, SCD Type 2, schema tests) · Apache Airflow
                Snowflake · AWS (S3, Redshift, Glue) · Azure (EventHub, Data Factory, Key Vault)
                Medallion Architecture · Dimensional Modeling

Backend         FastAPI · Flask · REST APIs · PostgreSQL · MySQL · Docker · GitHub Actions CI/CD

Visualization   Streamlit · Plotly · Folium · Tableau · Power BI

Languages       Python (pandas, NumPy, PySpark) · SQL · R · Bash

AI Engineer — Vision Systems & Intelligence Labs (Jan 2025 – Present) Multimodal healthcare AI on 119,533 MIMIC-CXR records. Fine-tuned medical LLMs under hardware constraints. Published at SPIE Medical Imaging.

Data Engineer — Center for Railways Information System (Dec 2021 – Jul 2023) ETL pipelines over 1M+ records. 45% query latency reduction via MySQL rewrites, composite indexing, and table partitioning. Eliminated 8 hrs/week manual reconciliation. Docker + GitHub Actions CI/CD cut deployment from 2–3 weeks to under 2 days.

AWS Certified Data Engineer – Associate (DEA-C01) — Amazon Web Services · June 2025

MS, Information Science – Machine Learning · University of Arizona · 3.89 GPA · 2023–2025