AI Medical Scheduling Agent — Multi Organization Voice & Chat Scheduling Platform (Deployed on AWS)

Outcome: Enterprise production voice AI and chat multi organization scheduling platform that turns natural patient conversations into validated physician routing and real appointment bookings.

• Built natural-language patient intake that converts conversational speech into structured medical categories, identity data, and scheduling inputs.
• Engineered deterministic physician routing, appointment matching, confirmation, and transactional booking with safeguards against malformed data and double booking.
• Used LangSmith tracing and benchmarking to evaluate AI behavior, then containerized and deployed the full platform to AWS with Docker and Nginx.

Private repository and documentation.

Vizor – Pharmacy Verification Computer Vision for Robotics

Outcome: Enterprise grade computer vision and sensor fusion platform for automated pharmacy verification in robotics and medical technology.

• Used a ConvNeXt deep learning model together with sensor data to verify medication trays, labels, and doses.
• Achieved 96.7% accuracy on the held-out test set, with inference latency under 100 milliseconds on NVIDIA GPUs and Apple MPS.
• Included a FastAPI inference service, fail closed safety logic, reproducible evaluation pipeline, model card, and dataset management.

Sophia AI - A Standalone Local LLM for Technical Interview Coaching (100% Offline, No Cloud Calls)

Outcome: Domain specific, fine tuned Large Language Model for technical interview coaching.

• Designed to act as an expert technical interviewer for AI and ML, backend, and MLOps roles.
• Built as a standalone 1.5B parameter neural network fine tuned with QLoRA on a proprietary Enterprise Knowledge Lake.
• Supports local inference on Apple Silicon or NVIDIA GPUs with zero external API dependencies.

MiahAI – Offline, Full-Stack AI Platform for Private Learning

Outcome: Production ready AI learning application designed to teach instead of only handing out quick answers.

• Provides context grounded in proven teaching methods shaped by decades of research on how people learn and retain knowledge.
• Designed to behave as a learning application, not a basic answer calculator.
• Built as a private, full-stack AI platform for learning workflows.

Offline Modular Reasoning Engine: Testing Miah LLM

Outcome: Client agnostic LLM reasoning service with server streaming gRPC.

• Built hot swappable model adapters for OpenAI, DeepSeek, Grok, Azure APIs, and a local AI Miah LLM.
• Supported local model execution via ONNX Runtime and TensorRT for LLMs with SQLite session memory.
• Shipped health and readiness endpoints, structured logging, unit tests, and a Python client.