AI/ML Engineer building production data & LLM systems.
I design and ship machine learning pipelines, RAG and agent systems, and the data infrastructure underneath them — with 6 years focused on taking models from notebook to production, reliably and at scale.
Experience
Microsoft
- Built end-to-end RAG pipeline on Azure AI Search (hybrid/semantic search) to parse 500 PDF invoices, serving grounded, zero-hallucination GPT-4o answers via FastAPI with cross-document mathematical aggregation.
- Built XGBoost churn model on 8K tenants using SHAP TreeExplainer for risk attribution; automated personalized GPT-4o retention outreach via Semantic Kernel for accounts with probability ≥ 0.60, scaling output to Cosmos DB and executing a segment-level Responsible AI fairness audit.
- Designed a 4-agent AutoGen pipeline (Researcher, Analyst, Writer, Compliance) on GPT-4o via Azure Functions, with a dedicated compliance-review stage and full audit-trail logging to Cosmos DB; diagnosed OpenAI rate-limiting as the cause of 7/20 pipeline failures, fixed it via retry backoff and request pacing to reach 100% completion (up from 65%), and verified the compliance gate by confirming it correctly rejects a deliberately non-compliant draft.
- Built FastAPI safety middleware layering Azure Content Safety and a custom prompt-injection blocklist to screen both user input and model output; iterated on a labeled evaluation harness (60 prompts) to raise recall on harmful content from 70% to 90% (via a zero-tolerance child-safety keyword check and recalibrated severity thresholds) while holding precision at 100% and false-positive rate at 0%.
University of Florida
- Built end-to-end lung nodule detection pipeline on LUNA16 dataset; trained a 3D CNN (AUC=0.88) outperforming a ViT counterpart, integrated SAM for zero-shot slice segmentation, and deployed via FastAPI and ONNX.
Samsung Research Institute
- Delivered 10+ production software releases serving 50M+ users at Samsung Research Institute; 99.5% crash-free rate and 40% bug reduction through rigorous testing and performance optimization
Featured Projects
Local Voice AI Chat
A fully local, privacy-preserving voice assistant. Speak into your microphone, your voice is transcribed on-device with MLX Whisper, and a local LLM served by Ollama answers you in a Streamlit chat UI — no audio, text, or API calls ever leave the machine.
On-Device Vision Pipeline
A four-stage computer vision pipeline: classify business charts (MobileNetV3, transfer learning), detect chart regions inside dashboards (YOLOv8, bounding boxes derived for free from layout), compress for on-device deployment (CoreML FP32 → FP16 → INT8 → 4-bit), and adapt to a new visual theme from just 160 images via a hand-rolled LoRA adapter.
Agentic Sales Pipeline
A 4-agent AutoGen pipeline (Researcher, Analyst, Writer, Compliance) that turns a company name into a compliance-approved, personalized outreach email, with full audit-trail logging to Cosmos DB and deployment as an Azure Function. Diagnosed OpenAI rate-limiting as the cause of 7/20 pipeline failures, fixed it via retry backoff and request pacing to reach 100% completion (up from 65%), and verified the compliance gate rejects a deliberately non-compliant draft.
Copilot Safety Middleware
Safety middleware for a Copilot-style chat API, screening input and output through Azure Content Safety and a custom blocklist. Found and fixed two real bugs (dead blocklist-checking code, miscalibrated severity thresholds), verified live against the API: recall improved from 70% to 90% while holding 100% precision and 0% false-positive rate.
Multi-Cloud LLM Platform
A provider-agnostic LLM routing layer spanning AWS Bedrock, OpenAI, and local Ollama models, instrumented with Prometheus metrics, wrapped in a LangGraph iterative research agent, and exposed to any MCP-compatible client (e.g. Claude Desktop) as callable tools over an MCP stdio server.
About
I came to ML through statistics and computer science — CS gave me the ability to build systems, statistics gave me the framework to reason about whether they actually work. I'm especially drawn to finance and healthcare: both are fields where "pretty good" isn't good enough, and where real constraints — regulation, interpretability, actual consequences for being wrong — force more careful engineering than the average ML project., Outside of work: I spend a lot of my own time reading and learning about how ML is actually applied in finance and healthcare — less the hype, more what's genuinely working in practice.
