I'm an AI Engineer.

Theodore Roosevelt Presidential Library

• FastAPI • WebSockets • Azure

Production multi-agent conversational AI for a permanent museum installation. A primary TR-voice agent answers visitors in-character, supported by classification, context tracking, and note-taking agents.

• Engineered real-time LLM orchestration over WebSockets handling kiosk connections, prompt assembly, context management, and low-latency streaming responses via Azure Speech
• Implemented layered guardrails: upstream input classification, downstream output inspection, and a fixed top-level prompt layer enforcing safety constraints across CMS-driven content
• Designed prompt assembly to keep responses on-character within strict latency budgets

The National Archives

• FastAPI • Azure AI Search • CLIP/BGE

Semantic retrieval and ranking system powering 30 interactive gallery stations in The American Story exhibit, surfacing relevant matches from a 2M-record curated set drawn from a 15M-record corpus.

• Engineered multimodal ingestion pipeline with LLM-based enrichment and CLIP/BGE embeddings to build the retrieval indexes
• Trained an engagement-based ranking model to filter the corpus and provide query-time signals
• Built offline filtering and a query-time ranking system using user-labeled feedback collected through a purpose-built evaluation interface

Identity-Conditioned Photo Platform

• Gemini • GPT • AWS • Terraform

Identity-conditioned image generation that places visitors into themed scenes — powering permanent installations at the Theodore Roosevelt Presidential Library and another museum, plus a conference-expo kiosk used to sell the capability.

• Built the expo kiosk as an async end-to-end product: mobile QR-loaded selfie form, generation backend, kiosk display, and an admin panel for switching between theme modes (TRPL, spy disguises, dinosaurs, and more)
• Added an LLM-driven theme-creation agent — visitors chat to define a new theme, the agent emits a structured prompt config, and admins review before publishing it to the kiosk
• Integrated Gemini and GPT Image behind a unified API with SFace-based identity-similarity evaluation for runtime quality gating, deployed via Terraform on AWS (Elastic Beanstalk, S3, CloudFront)

Architecture Visualizations

• Python • ComfyUI • Gemini

Image-generation service for an architecture museum that produces concept views of buildings using a two-tier hybrid pipeline.

• Local diffusion models for fast, low-cost previews
• Gemini for the final polished hero output

Editorial Sound Bite Surfacing

• Whisper • GPT batch • Claude

LLM-driven workflow that surfaces candidate sound bites from interview transcripts for editorial review — decomposed so the producing team could own the final matching step themselves in Claude.

• Whisper transcribes interviews; a batch GPT pass tags each utterance with structured metadata
• Final brief-to-utterance matching handed off to producers as a Claude prompt — putting frontier AI directly into the editorial workflow

Content Safety Classifier

• .NET 8 • ONNX • BERT/RoBERTa

BERT/RoBERTa-based profanity classifiers trained as a content-safety guardrail layer, optimized for low-latency inference via ONNX in a .NET 8 runtime.

• Fine-tuned transformer classifiers on profanity detection
• Exported and optimized for ONNX runtime in .NET 8 inference services

Width by Height

• Flutter • Swift • Firebase

iPad-first app for designing artwork arrangements. Import your artwork or browse The Met, photograph your space, drag and drop, and preview arrangements in your room with AR.

• Built the Flutter app with Firebase integration
• Implemented ML models from current research to predict room geometry from RGB images
• Used ARKit to detect, perspective-correct, and predict size of scanned artwork