Dev Narayan C. B.

Architected and deployed a multi-sensor welfare-monitoring research framework designed for non-intrusive Activities of Daily Living (ADL) tracking in elderly-care environments.

• Multimodal Sensor Fusion: Built a synchronized data acquisition system integrating TI AWR6843AOP mmWave FMCW radar, Sensfloor pressure-sensitive flooring, SOMNOtouch physiological wearable sensors, and Captury markerless motion capture across 13 distinct human activity classes.
• Rigorous Experiment Pipeline: Designed the experiment pipeline end-to-end: standardized dataset acquisition across three rooms in a nursing-home simulation lab, session-based train/test splits to eliminate temporal/spatial leakage, and per-class decision threshold tuning to mitigate extreme class imbalance.
• Real-Time SSL on Consumer GPUs: Extended a self-supervised detection architecture (Barlow Twins-based, from 2023 publications) into a real-time inference pipeline running on consumer-grade GPUs without losing detection mAP.
• Data Integrity & Bug Diagnosis: Diagnosed and resolved severe data-integrity anomalies in raw sensor captures (split binary formats, packet loss under heavy USB bus load, session contamination) that were silently corrupting downstream training models.

First AI hire; led development of an agentic RAG platform for legal professionals from prototype to 200+ paying customers. Spearheaded machine learning architecture, document digitisation pipelines, and low-latency microservices.

• Agentic Legal RAG: Fine-tuned LLMs and built agentic RAG pipelines over Indian case-law databases and court judgments for case-reference, argument generation, and statutory citation validation.
• Multilingual Indic OCR Engine: Engineered a deep document-understanding engine capable of digitizing degraded physical legal petitions, court affidavits, and regional language documents across Indic scripts with automated statutory reference linking.
• High-Throughput Distributed Backend: Designed and deployed backend microservices using FastAPI and Node.js for real-time document indexing, SSE token streaming, and collaborative document markup.
• Vector Indexing & Cost Optimization: Optimized Milvus vector clusters with semantic caching and chunk re-ranking, reducing LLM token consumption by over 60% and trimming median retrieval latency to sub-second speeds.

Developed high-accuracy document intelligence pipelines automating customs shipping declarations and complex invoice processing.

• Customs IDP Pipeline: Built an Intelligent Document Processing engine extracting structured schema from diverse multilingual printed customs bills and commercial invoices.
• Multi-Engine OCR & Noise Reduction: Integrated Azure OCR, Gemini, and GPT with automated pre-processing filters, minimizing character confusion errors and optimizing API throughput.
• Automated Profile Generator: Developed an automated CV synthesis engine structuring raw professional biographies into standardized, ATS-compliant formats.