BHARAT DHYANI

TRC Companies – LLM Research & Production Deployment

• Architected and deployed production-grade LLM applications using OpenAI GPT-4, Llama 2, and Mistral models with Retrieval Augmented Generation (RAG) for processing contracts, RFPs, grants, and financial documents
• Built agentic pipeline leveraging LangGraph, CrewAI, and AutoGen to autonomously extract relevant information from RFPs and generate tailored proposal documents with 70% reduction in manual effort
• Productionized multi-agent systems using FastAPI backend with asynchronous processing, handling 1000+ concurrent requests with
• Implemented vector database solutions (FAISS, ChromaDB) for efficient document retrieval with semantic search capabilities, improving retrieval accuracy.
• Developed automated entity extraction pipeline using NER and custom fine-tuned models to populate structured documents from unstructured PDFs.
• Built SQL Coder LLM (defog-ai) powered natural language to SQL converter, enabling non-technical users to query databases with 90% query accuracy.
• Engineered LLM resource finder application with context-aware search, reducing resource discovery time by 60%
• Fine-tuned LLMs using advanced quantization techniques (GGML, LoRA, QLoRA, LofTQ) to deploy high-precision models on limited GPU resources with 4-bit quantization
• Implemented comprehensive prompt engineering framework with security guardrails and compliance validation.
• Deployed scalable inference infrastructure on AWS (EC2, Lambda, API Gateway) supporting 10K+ daily API calls with auto-scaling capabilities

Helix – Recommendation System for HNI Clients

• Built hybrid recommendation engine combining content-based filtering and collaborative filtering using matrix factorization (SVD, ALS) for personalized investment recommendations
• Implemented advanced feature engineering pipeline extracting 50+ behavioral and demographic features from client data
• Developed rule-based recommendation layer using K-means clustering on client profiles, improving recommendation relevance by 42%
• Achieved 78% recommendation acceptance rate through A/B testing and continuous model refinement
• Additional Projects & Initiatives

MCC Healthcare:

• Built voice-to-text transcription system for patient-provider conversations using Whisper API for automated claim processing, reducing processing time by 55%
• Document Intelligence: Developed PDF table extraction system using Camelot with NLP-based information extraction from 10-K reports for automated financial analysis
• Received multiple recognition awards for implementing cutting-edge LLM solutions and conducting organization-wide training sessions on Gen AI
• Developed custom scripts for cleaning messy datasets, ensuring high-quality inputs into the analytic process.
• Enhanced decision-making capabilities by creating interactive dashboards and visualizations.

Canaery – Neural Data Processing for Olfaction Detection

• Developed encoder-decoder deep learning architecture for olfaction (smell) detection using neural spike data, achieving 85% classification accuracy for security applications
• Engineered end-to-end preprocessing pipeline on AWS SageMaker for processing complex neural data formats with automated quality validation
• Built real-time neural data streaming and analysis system with spike event detection and visualization dashboards
• Architected serverless data pipeline using AWS Lambda, API Gateway, and S3 for automated ETL processes handling 100GB+ daily neural recordings
• Reduced data processing time by 70% through optimized batch processing and parallel computing on EC2 spot instances

• Architected and deployed production-ready ANPR (Automatic Number Plate Recognition) system achieving >95% accuracy on edge devices (NVIDIA Xavier, Jetson Nano) with
• Optimized deep learning models using TensorRT and knowledge distillation, achieving 9-10% mAP improvement while reducing model size by 60%
• Built real-time video analytics pipeline processing 30 FPS live streams with YOLOv5 and DeepSORT tracking for vehicle monitoring
• Implemented synthetic data generation pipeline for license plates addressing class imbalance, improving model mAP by 15%
• Developed MLOps infrastructure including A/B testing framework, automated training pipelines, and auto-labeling tools reducing annotation time by 80%
• Engineered car logo detection and tracking system using YOLOv5 with centroid tracking, deployed in production serving 50K+ daily inferences

• Built intelligent chatbot for CRM applications using NLP and intent classification, reducing support tickets from 1000+ to
• Developed fraud detection ML system analyzing historical user data, successfully identifying 7% fraudulent activities and implementing preventive measures
• Created Power BI dashboards for complaint pattern analysis, enabling data-driven decision making for support optimization
• Engineered automation tools for module generation reducing developer workload by 40%