Ankur Das

Summary

• Contributed to the data acquisition pipeline for time series sensor data, assisting in edge optimization with TFLite and TFLite-Micro for deployment on resource-constrained devices. Enabled data gathering from multiple clouds and direct machine connections using IoT protocols like MQTT. • Crafted an edge-optimized real-time video processing system for dynamic color correction using an autoencoder-based model, optimized with ONNX. Achieved 84% accuracy and 33 fps performance on resource-constrained edge devices, ensuring seamless video enhancement. • Engineered a computer vision system to identify faulty semiconductor chips from images using a robust training and inference pipeline. Leveraged data augmentation, CNNs, and GANs to enhance model performance, addressing signif- icant data imbalance through advanced augmentation and ensemble learning. Achieved 86% precision and improved faulty chip classification accuracy by over 50%. • Implemented a real-time security surveillance system using object detection on video streams. Fine-tuned pretrained models to identify target objects in specific environments and deployed them on an Intel NUC edge device using OpenVINO which achieved 83% mAP and 110ms latency. • Python, C++, Scikit-learn, Tensorflow, Keras, OpenVINO, Git

Summary

• Spearheaded the development of a no-code/low-code Agentic Service to orchestrate complex workflows, APIs, and multi-agent systems. Devised a modular framework supporting both proprietary and open-source LLMs and vector databases. Incorporated semantic similarity search and conversational memory to enhance response relevance. Created an intuitive system where agents can be built with just a few clicks, reducing setup time by over 90%. • Conceptualized and deployed a production-grade Retrieval-Augmented Generation (RAG) system capable of ingesting diverse document types (PDF, TXT, MDX) for natural language querying over internal knowledge bases. Seamlessly embedded into an agentic platform, utilizing semantic similarity search and optimized chunking strategies to maximize retrieval effectiveness. Achieved 92–97% Recall@10 for context relevance and 95-98% groundedness and answer relevance on test data. • User formulated an automated service for hierarchical taxonomy generation using LLMs, combining prompt-based workflows with an iterative refinement loop to improve structure and tagging accuracy. Reduced taxonomy generation time from days to hours and enabled production-ready tagging pipelines, allowing domain experts to customize solutions by simply updating prompts. The system is evaluated on three benchmark datasets and depending on attributes, achieved 90-98% accuracy. • Built a supervised learning framework for automated text data tagging, ground truth validation, and evaluation, supporting models like LLM-powered regex, BERT, T5, and XGBoost. Constructed an end-to-end pipeline with human-in-the- loop QA, automating over 95% of training workflows. Achieved 92–98% accuracy across three benchmark datasets, demonstrating broad applicability and high generalization. • Orchestrated a dental diagnosis system combining object detection and image classification to assess the severity of dental conditions. Applied ensemble learning to improve diagnostic accuracy and prediction robustness. Integrated a recommendation module to suggest the most suitable dental center based on diagnosis and patient location, enhancing timely and accessible treatment. Achieved a 13% increase in precision and improved model explainability, boosting trust in clinical decision support. • Python, PyTorch, Scikit-learn, Huggingface, LangChain, Ollama, ChromaDB, ComsosDB, REST API, Microsoft Azure, Git, Docker, Kubernetes

Summary

• Built a real-time sentiment analysis system for phone calls using MFCC for audio feature extraction and CNNs for classification. Designed for customer service applications, achieving 68% accuracy in emotion prediction. • Python, Keras

Summary

• Constructed a question-answering system to generate questions and answers from essays. Implemented text sum- marization and splitting algorithms to identify relevant paragraphs. Utilized two T5 models for question and answer generation, achieving around 90% precision for questions and 86% precision for answers. • Python, PyTorch, FastAPI, Git