02. spec // projects
Automates material classification on constrained hardware so sorting decisions can happen at the edge, on the robot, without a cloud round-trip. Delivered 0.87 mAP at 30 FPS by quantizing and deploying a ViT + DETR pipeline through ONNX on Raspberry Pi.
Detects eye closure, yawning, and head-pose drift early enough for an in-cabin edge device to intervene in real time. Reached 95% accuracy with sub-50 ms latency and stable 30 FPS inference on Raspberry Pi hardware.
Localizes drywall defects in real-world inspection workflows without a fully supervised dataset for every defect class, the same data bottleneck robots hit in the field. Built a prompt-driven segmentation pipeline with SAM and Grounding DINO for zero-shot defect localization.
Lets operators ask free-form questions about fleet-efficiency data and get narrated, number-backed answers instead of hand-writing SQL. Built a guarded text-to-SQL agent over DuckDB with an IsolationForest anomaly detector (0.98 ROC-AUC) in one Streamlit app.
Classifies music genres by combining spatial Mel-spectrogram features with temporal sequence reasoning over raw audio signals. Built a CNN + BiLSTM pipeline over MFCC and spectrogram features with roughly 92% accuracy on GTZAN.
Coordinates research, drafting, critique, image generation, and publishing through one LangGraph task orchestration system. Adds autonomous routing, LLM refinement loops, checkpointed state, revision controls, and real-time Streamlit monitoring.
Finds relevant researchers and turns noisy public web data into structured outreach targets for academic collaboration. Combined TF-IDF, cosine similarity, multi-threaded scraping, Llama 3 agent flows, and FastAPI outreach automation.