I ship production AI systems.
5+ years building software, focused on getting LLM systems into production on Claude & OpenAI. I design RAG / GraphRAG pipelines, LLM agents (LangGraph, MCP, tool use), fine-tuned models, and FastAPI backends — owning deployment, CI/CD and monitoring end to end.
I build AI that ships, not AI that demos. Most LLM projects stall at the prototype — mine go to production, get monitored, and keep working when real users show up.
I've gone from co-founding an edtech startup, to architecting enterprise GenAI serving 50K+ users, to building computer-vision pipelines for the power grid. The throughline: I own the whole problem — model, backend, deployment, and the metrics that prove it worked.
Demos are easy; production is the job. I deliver deployed, observable, maintainable systems — not notebooks.
Every claim ties to a number — latency, accuracy, coverage, cost. If it can't be measured, it isn't done.
You'll always know what I'm building and why — clear trade-offs, honest status, no black boxes.
Hackathon wins, published research & open source
Built a novel multi-stage synthetic data pipeline for complex relational reasoning using Python and PyTorch. Implemented Guided Reinforcement Policy Optimization (GRPO) with custom JSON rewards for logical accuracy. Fine-tuned Llama 3, Qwen 3, and Phi-4 with LoRA, creating separate Q/A model pairs, and identified and mitigated a major LLM reasoning flaw (alphabetical sorting bias).
📖 Read the technical deep dive →Shao, A., Carradini, S., Harikrishnan, R. S., et al. (2026). "Currents of Inquiry: Insights From Two Years of Real-World AI-Learner Water Conversations." Proceedings of LERN 2026, pp. 193–201. Co-authored peer-reviewed analysis of two years of conversational data from the Arizona WaterBot project — how real users interact with AI for information retrieval.
Developed a wildfire simulation environment enabling reinforcement learning research and agent training. Collaborated with Meta / PyTorch teams to enhance simulation capabilities, and wrote comprehensive technical documentation and unit tests that improved adoption and developer workflows.
5+ years building scalable systems & intelligent applications
Freelance & contract · fixed-scope or ongoing
Chatbots and search that cite your data. Chunking, embeddings, vector + graph retrieval, and guardrails that fight hallucination — built to scale.
Agentic workflows that use tools and complete multi-step tasks reliably. LangGraph orchestration, MCP integrations, structured output.
Custom models for your domain. LoRA/QLoRA and GRPO on Llama 3, Qwen 3, Phi-4, backed by synthetic-data pipelines and proper eval.
Production API layers: async endpoints, streaming responses, Pydantic validation, auth, caching, and metrics — Dockerized and deploy-ready.
Off the laptop and into production: AWS / Azure / GCP, CI/CD, containerization, model versioning, A/B testing, and monitoring.
Detection and segmentation for real imagery — drone, ground-level, street-view. Model benchmarking, data scraping, reproducible training.
A multi-stage synthetic data generation pipeline for the Meta AI / AMD / PyTorch Synthetic Data AI Agents Challenge. A relationship-aware system that generates high-quality question-answer pairs for training LLMs on reasoning tasks involving character relationships and seating arrangements. Implemented GRPO with custom reward functions, structured JSON output enforcement, and identified critical LLM vulnerabilities like alphabetical sorting bias. Fine-tuned Llama 3 70B, Qwen3-Next 80B MoE, and Phi-4 using LoRA, SFT, and GRPO.
📖 Read the technical deep dive →
Production-ready GraphRAG app: FastAPI backend, Neo4j graph DB, and OpenAI GPT-4o for query processing. 40+ API endpoints, Redis caching (200× faster cached responses), Prometheus monitoring, 70%+ test coverage, Docker + CI/CD.
Reached the Final Four with a reinforcement-learning system using Llama 3.2 with LoRA fine-tuning. Trained on 16,000+ demonstrations using GRPO and behavior cloning on AMD MI100 GPU. Deployed to Hugging Face Spaces with optimized inference.
An AI tutor that answers math, physics & science questions and generates interactive React visualizations in real time using the Claude API. Claude returns explanation + executable component code, safely rendered inside a sandboxed iframe.
Large-scale graph data-processing pipeline using Neo4j, Terraform, AWS, Docker, Kubernetes, and Kafka. Processed the NYC Yellow Cab dataset into graph nodes/relationships, implementing PageRank and BFS with Neo4j Graph Data Science.
Collaborative AI trip planner where groups plan via natural-language chat. A LangGraph agent extracts trip memory with confidence scoring, versions plans with rollback, pulls live flight data (Amadeus), and streams updates over Server-Sent Events.
Deep-learning model to classify AI-generated vs. real images using ResNet50v2 and EfficientNetV2. Preprocessed CFAKE & ArtiFact datasets with augmentation, performed hyperparameter tuning, and evaluated with AUC, accuracy, precision, recall, and F1.
Hands-on, in production — not buzzwords
Academic background & coursework
Technical deep dives & project stories
How a hackathon prototype became a production-ready RL environment in Meta's PyTorch OpenEnv framework — physics-informed fire simulation, AI agent training, and an interactive human-vs-AI game.
Read more →Designing and deploying a GraphRAG-powered multilingual chatbot serving 50,000+ users with real-time water-resource information across Arizona communities.
Read more →Available for freelance & contract work — RAG & GraphRAG, LLM agents, fine-tuning, and Python/FastAPI backends. Also open to full-time ML & software engineering roles. Tell me what you're building.