Bhavneet Singh

• Architected a Graph-RAG multi-agent system enabling distributed inference over 150K+ documents, supporting 40+ stakeholders. • Built multi-agent orchestration (LangGraph + CrewAI + MCP + Ray) with policy routing, improving decision accuracy by 26%. • Implemented hybrid retrieval (MiniLM-L6-v2 + ColBERT) with feedback loops, improving retrieval precision@k by 31%. • Optimized FAISS, Milvus, and PostgreSQL vector stores to index 3M+ legal docs with sub-second retrieval latency. • Deployed LLM inference on vLLM + Triton, enabling multi-tenant serving with predictable latency/memory-efficient utilization. • Designed LLM microservices (FastAPI + ONNX + CUDA graphs) achieving <400ms inference under concurrent production load. • Applied VLMs (CLIP, BLIP-2) with hybrid retrieval and contextual reasoning, improving real-time inference accuracy by 18%. • Fine-tuned LLMs/vision models using LoRA-based tuning and quantization, boosting accuracy by 24% and reducing size by 48%. • Developed distributed ML pipelines (PySpark + Spark + PyTorch) over multi-node clusters, improving throughput by 65%. • Built real-time ingestion and embedding pipelines with PII filtering/schema validation, accelerating ETL throughput by 52%. • Integrated LangSmith + MLflow + W&B observability, reducing debugging cycles by 40% and standardizing RAG quality metrics. • Automated drift detection, retraining loops, and lineage tracking via Databricks, reducing performance degradation by 33%. • Implemented token tracking, prompt versioning, and cost monitoring to control LLM spend and predictable production scaling.

Teaching Assistant/Demonstrator/Lab Monitor -Intro to Software Engineering and Data Science • Supervising Labs on Software Engineering and Android concepts in Java, grading Assignments, Proctoring Exams, and Conducting Office Hours for doubt resolution. • Mentored labs in Data Science, guiding students in Python programming, and automated ML using pandas and scikit-learn.

• Engineered an agentic-AI benchmarking framework that autonomously identifies and corrects regression imbalance across 34 real-world datasets spanning cybersecurity, climate, housing, and healthcare. • Integrated LangChain and AutoGen agents for adaptive data relevance estimation and self-driven model retraining workflows. • Developed autonomous analysis agents leveraging φ-weighted metrics (SERA, SERT, φ-RMSE/MAE) and SMOGN-based oversampling, reducing rare-event prediction error by up to 3×. • Designed CrewAI-based multi-agent pipelines coordinating preprocessing, hyperparameter tuning, and diagnostics, improving runtime efficiency by 45%. • Implemented AI reasoning modules using KDE, GMM, and spectral density estimators for self-assessment of model reliability in high-impact regions. • Created bias-detection and interpretability agents leveraging HuggingFace Transformers, uncovering 60% of previously missed rare outcomes. • Automated research reporting through an LLM-powered summarization agent (LangChain + LaTeX), cutting documentation time by 70%. • Released as a PyPI-versioned, CI/CD-enabled package with reproducible configs and agent orchestration templates, adopted by multiple AI research labs.

• Engineered agentic AI pipelines for context-aware orchestration across varied multi-step workflows supporting 10+ clients. • Developed RAG workflows for document search using HNSW-based ANN indexing with Pinecone, improving retrieval precision by 22%. • Built end-to-end data pipelines using Python, AWS, and MLflow for transformer-based models, improving accuracy by 12%. • Implemented a batch preprocessing pipeline for large document inputs, reducing parsing time by 8% and improving backend performance. • Reduced inference time by 30% via a custom FFT-based model, optimized for real-time text categorization. • Processed 500K+ document requests with REST APIs on AWS EC2, sustaining <200ms latency. • Developed React.js dashboard for admins, simplifying monitoring & results visualization. • Integrated CloudWatch monitoring for ML endpoints, cutting debugging time by 60% and improving uptime. • Automated ML infrastructure CI/CD with GitHub Actions and Terraform, supporting scalable deployment across client systems.

• Improved data retrieval speed by 35% via optimized SQL joins, indexed queries, and Redis caching in the Express.js service layer. • Optimized backend services with FastAPI and MongoDB, enabling fast retrieval of dynamic datasets and improving data access performance by 25%. • Reduced dashboard load latency by 20% using streaming visualizations with React.js, D3.js, and real-time backend event handling via WebSockets. • Automated ML training workflows on AWS SageMaker for 2.5M+ records, enabling smooth backend integration and scalability. • Performed statistical analysis and dimensionality reduction (PCA, Bayesian inference) to optimize ML model inputs. • Delivered containerized Python APIs for 3 production ML models via MongoDB and Docker, reducing deployment friction by 40%.

Resolved over 1000 doubts, performed debugging on various algorithmic problems, guided more than 120 students for projects and assignments.

• Automated ingestion and preprocessing of large-scale geospatial climate datasets via Google Earth Engine, transforming and cleaning data using PostgreSQL and pandas for ML-ready pipelines. • Built and trained hybrid CNN–LSTM models in TensorFlow to classify crop outcomes from time-series climate data, achieving 91% precision and 90.1% recall. • Accelerated model convergence by 30% through hyperparameter optimization (batch size, learning rate) on AWS EC2 GPU instances. • Implemented robust model diagnostics and monitoring with TensorBoard, AUC–ROC, and confusion matrices to ensure sustained accuracy across training cycles. • Deployed trained deep learning models through Flask REST APIs with integrated logging and exception handling, enabling research teams to perform live forecasting and model validation.