.png&w=256&q=75)
š¤ Latest Submissions
AI Smart Warehouse Safety & Audit System
Core Architectural Components Edge Vision Processing Layer: Automates spatial downsampling across massive frame datasets, sampling sequential video frames dynamically to dramatically optimize hardware compute efficiency. Fine-Tuned Multimodal Inference Pipeline: Features a specialized LoRA adapter running natively on PyTorch ROCm infrastructure. The model dynamically interprets visual layout complexities, localized machinery activity, and granular safety variables (such as personnel tracking and PPE compliance metrics). Rugged Normalization Engine: A custom, robust post-inference parsing layer that scrubs syntactical anomalies, cleans inline VLM commentary via regex parsing, and unifies variant output schemas in flight to eliminate unparsed data losses. Transactional Data Ledger: Streams structured audit insights concurrently into high-fidelity JSON time-series ledgers and structured CSV alert records to maintain an immutable compliance log. Live Operational Dashboard: A responsive Streamlit web UI that ingests the transactional ledgers in real time, converting raw multimodal inferences into actionable risk heatmaps, key metrics, and immediate safety indicators for warehouse operators. Key Technical Achievements ROCm Hardware Optimization: Fully utilized local AMD compute kernels, managing stable multi-gigabyte VRAM bounds while keeping core execution times uniform across high-volume batches. Dynamic Content Extraction: Built-in resilience to edge-case model hallucinations and syntax shifts, turning brittle string parsing into a rugged, enterprise-grade validation layer. I generated the cover image with prompts tailored to my project in Gemini.
13 Jul 2026
