Nitisha Adusumilli

Architected LangChain RAG pipelines with FAISS, improving grounded-answer relevance by 16% using embedding generation, reranked retrieval, source attribution, and structured evaluation across product and marketing datasets. Built LangGraph agents to orchestrate tool calls for dataset tracking, experiment logging, and evaluation reporting with structured state management and resilient REST integrations. Fine-tuned Meta Llama models with Hugging Face Transformers and DeepSpeed ZeRO-2, increasing distributed training throughput by 21% while controlling memory across multi-GPU experiments. Implemented KNN identity resolution with FAISS similarity scoring and calibrated thresholds using precision-recall analysis plus semantic enrichment to produce reusable downstream ML feature vectors. Productionized ML workflows using PostgreSQL, React, Docker, and CI/CD, reducing experiment-to-reporting turnaround by 18 hours.

Benchmarked streaming speech-to-text models across accented, noisy, and telephony audio using Python and PyTorch, improving word-error-rate performance by 14% on targeted evaluation datasets. Tuned audio chunking, endpointing, and streaming inference to cut median transcript latency by 18% and preprocessing overhead by 11% during multi-turn conversational voice tests. Built WebSocket test clients to replay live audio, capture partial transcripts, and isolate endpoint failures, reducing speech-pipeline regression triage time by 24% across production releases.

Trained TensorFlow classifiers on sensor-derived signals for an on-device inference prototype, improving validation accuracy by 12% through augmentation, feature refinement, and class-balancing experiments. Profiled model latency and memory for constrained device targets and applied post-training quantization and graph optimization to reduce inference time by 17% without destabilizing production predictions. Investigated misclassifications using NumPy and Pandas to trace noisy sensor windows, label imbalance, and preprocessing drift, then refined feature pipelines through systematic embedded-device validation. Calibrated anomaly-detection thresholds against device telemetry distributions to lower false positives by 15% while preserving sensitivity to abnormal patterns during hardware-in-loop evaluations. Standardized experiment configurations, checkpoints, and evaluation reports to improve model reproduction turnaround by 20% across ML and embedded-software validation handoffs.