Architected an AWS data lake on Apache Iceberg, Athena, and Kinesis Firehose to ingest and query per-frame CV telemetry across a 19-machine fleet at 5.4M frames/week. Designed a self-correcting MLOps pipeline that runs weekly diagnostics via embedding drift detection, confidence monitoring, and label mismatch analysis, classifies root cause, and triggers targeted finetuning on RunPod Serverless with automated backtesting and rollback. Built the entire production CV stack for real-time retail product tracking, spanning detection, edge inference, and a zero-shot auto-labeling pipeline (MediaPipe, YOLO-World, SAM2.1, Kalman filtering, BoT-SORT) at 95%+ box accuracy and a 7 hour train-to-production cycle.

Built end-to-end active learning pipelines in PyTorch that reached 90% of baseline accuracy using 5% of the original training data through uncertainty-based sampling. Enhanced feature-space sampling by replacing PCA+clustering with ResNet-18 embeddings for 2.5x greater accuracy improvement over the random baseline. Implemented custom training loops, data loaders, and automated retraining cycles.

Deployed C++ CNN applications for real-time object sorting, OCR, and anomaly detection on IoT hardware. Rewrote the C++ codebase with scheduling and parallel operations for a 6x speedup to real-time image processing. Implemented 2D and 3D computer vision algorithms such as convolutional edge finding and point cloud merging.

Researched prompting and agentic methods for LLMs to solve dynamic reasoning problems by formulating Markov Decision Processes and generating code solutions. Designed a chain-of-thought pipeline with LangChain that improved MDP formulation accuracy by 60%, and constructed RAG pipelines with LlamaIndex to extract 5000 datapoints from PDF textbooks; contributed to evaluation and prompt engineering for an agentic pipeline across GPT-4o, DeepSeek-R1, and Llama 3.3-70B.

ICLR 2025: Co-author on two accepted workshop papers on quantizing Protein Language Models (PLMs) and novel peptide generation from scarce data. Fine-tuned ProteinBERT, ESM-2, ProstT5, and ProLLaMA to predict protein function, structure, and stability. Researched quantization techniques for LLMs, leveraging LoRA and QLoRA fine-tuning to achieve 7:1 model compression while preserving 90% of validation accuracy. Developed flexible PyTorch pipeline for training HuggingFace models across generative and classification tasks. Built benchmark tool to search hyperparameters and visualize metrics, aiding performance evaluations

Developed Angular website in TypeScript to facilitate client access to Azure backend through microservice APIs. Created novel accessibility testing tool for Angular and integrated it with existing CI/CD pipeline. Conducted comprehensive unit tests on existing codebase using Jasmine and Karma.