<ul><li>AI‑native software manufacturing model: autonomous, agent‑driven SDLC with dynamic skill orchestration, AI‑human collaboration, and governed workflows for continuous improvement.</li></ul><p></p><ul><li> Architected a framework-agnostic observability and evaluation platform for enterprise agentic AI systems, enabling systematic analysis of agent behavior and performance.</li><li> Designed OpenTelemetry-based distributed tracing to capture end-to-end agent execution flows and support cross-SDK benchmarking across LangGraph, Google ADK, and Semantic Kernel.</li><li> Built modular evaluation pipelines integrating MLflow, LangFuse, and DeepEval for systematic assessment of agent quality, performance, and reliability.</li><li> Led security architecture and threat-modeling initiatives for enterprise agentic AI deployments.</li></ul><p></p><ul><li> Researching a hybrid approach to root-cause analysis for agentic systems that combines execution traces, logs, metrics, and domain context with algorithmic dependency-graph construction and LLM-based reasoning.</li><li> Exploring methods to localize failures from agent execution graphs, leverage LLMs for deeper root-cause investigation and remediation recommendations, and establish reliable evaluation strategies for validating proposed fixes.</li><li> Developing a proof-of-concept for a closed-loop diagnose–fix–re-evaluate framework, with ongoing research into fault localization, causal attribution, fix validation, and regression evaluation for agentic systems.</li></ul><p></p><ul><li> Partnering with an NGO to design and develop an AI-powered conversational platform for Sanskrit learning, leveraging language models to enable interactive practice, contextual dialogue, and adaptive learning experiences.</li></ul><p></p><ul><li> Designed a persona-driven synthetic data framework generating ~15M structured training data.</li><li> Distilled T5-small fom T5-base for conversational print-settings extraction, achieving 97% JSON exact-match accuracy.</li><li> Quantized the distilled model for resource-constrained edge deployment, optimizing memory footprint and inference efficiency.</li><li> Researched continual learning for implicit print-settings inference.</li><li> Built lightweight browser-based classifier (MobileNetV3, quantized, TensorFlow.js).</li></ul><p></p><ul><li> Led architecture of enterprise Excel layout-optimization plugin for print workflows.</li><li> Designed BFS-based table detection algorithm overcoming metadata limitations.</li><li> Delivered production-grade module integrated into enterprise systems.</li><li> Developed ML solutions for 3D printer analytics (anomaly detection, quality prediction, root cause analysis).</li><li> Designed LLM-based analytics frameworks and multimodal speech + NLP assessment systems.</li><li> Explored SLM experimentation (BitNet) for future lightweight model architectures.</li></ul>
