A multi-agent LLM system automating medical coding inside the Revenue Cycle Management pipeline. Each module (CPT, ICD, Modifier, NCCI Edits, LCD/NCD, MIPS) is handled by a dedicated agent with its own prompt strategy and retrieval flow.

• Architecture: LangChain + LangGraph multi-agent graph with agent-specific prompts.
• Retrieval: Semantic search over CPT & ICD via Vector DB (Pinecone / Qdrant) and PostgreSQL.
• Knowledge: LCD/NCD article matching, NCCI edits, and MIPS measures as retrievable sources.
• Fine-tuning: LLaMA 3.3 fine-tuned with LoRA for extraction & diagnosis prediction.
• Accuracy: CPT 88% · Modifier 86% · MIPS 100% · ICD 70%.
• Impact: Reduced manual processing time by 40%.

An end-to-end AI platform for automated 2D floorplan generation. Built on an enhanced DeepLayout framework trained on the RPLAN dataset, with a four-module sequential neural pipeline (Living → Continue → Wall → Location) and a vectorization engine that turns pixel outputs into professional vector floorplans. A Flask web app renders rooms with realistic textures.

• Up to 99.5% validation accuracy across the four modules.
• ~8 seconds per design vs ~45 min of manual MATLAB drafting (300× faster).
• 8/10 architects in user study would use the output as a starting point.
• Scope: single-story residential, 60–120 m². Supports 2/3/4-room layouts.

A two-stage CNN pipeline deciding whether an upper-limb X-ray shows a fracture. Stage 1 — a ResNet-50 body-part router picks Hand / Shoulder / Elbow. Stage 2 — a part-specialist ResNet-50 outputs fractured / normal. Ships as a CustomTkinter desktop GUI plus an optional Flask web demo.

• Trained on the MURA dataset (Stanford ML Group) · 20,335 musculoskeletal radiographs.
• Data augmentation (horizontal flips) + 72/18/10 train/val/test split.
• Adam optimizer (lr 1e-4) with early stopping.
• Body-part accuracy: 100% · Fracture detection accuracy: 87.9%.

Benchmarked four configurations on the CUB-200-2011 dataset (200 species, stratified 70/15/15 split): ResNet-50 baseline, + bbox crop, + MixUp/CutMix, and EfficientNet-B4 trained in two phases (warm-up + full fine-tune with label smoothing). Grad-CAM confirms the model attends to bird body, head and wing — not background.

• ResNet-50 baseline → 82.18% Val Top-1 / 95.93% Top-5.
• + bbox + MixUp/CutMix → 96.55% Test Top-5 (best ResNet variant).
• EfficientNet-B4 → strongest backbone, best overall.
• Reproducible pipeline: scripts for data, training, eval, Grad-CAM and checkpointing.

An AI voice system that places outbound calls to provider offices, navigates phone menus, speaks with staff, and records whether requested documents were received, need a resend, or require follow-up. Built as a production API with live call handling, structured outcomes, and secure result delivery.

• Architecture: Backend API + real-time voice agents. Separate paths for automated menus vs. live conversations.
• Speech: Cloud telephony with STT and TTS tuned for phone conversations.
• Outcomes: Transcript → structured result, artifacts stored, client notified post-call.
• Platform: Multi-tenant API, auth, queued bulk calling, cloud deployment.