Building agentic AI systems that reason, retrieve, and act.
I'm Pranesh Kannan — an AI Engineer and Data Scientist currently at IIT Madras, designing production LLM pipelines, RAG systems, and multi-agent workflows for enterprise supply chains and data platforms.
Designing systems that reason for themselves.
I'm an AI Engineer and Data Scientist working on agentic AI at IIT Madras — designing orchestration logic, tool-calling with ReAct reasoning, and RAG pipelines that turn unstructured enterprise data into grounded, autonomous decisions for supply chain and data engineering platforms.
What pulls me into this field is the moment an agent stops following a fixed script and starts genuinely reasoning — deciding which tool to call, when to pull in more context, when it has enough to act on its own. I spend a lot of my time reading agent-architecture papers, stress-testing new orchestration frameworks, and building up MLOps discipline (MLflow, Docker, Kubernetes) along the way — aiming for systems that are not just clever in a demo, but reliable enough to trust with real decisions.
Selected work.
Seven production and research projects spanning agentic orchestration, RAG, computer vision, and document intelligence — across IIT Madras, DRDO‑CVRDE, and ELAIT.
Supply Chain Decision Agent
An LLM-powered, agent-based model for autonomous supply chain decision-making — combining orchestration logic, MCP tool-calling, and ReAct reasoning to plan and execute multi-step workflows without manual intervention.
Agentic Data Product Layer
A modular pipeline that automates data preprocessing, annotation, ETL, and enrichment on Flink/Kafka, grounded by Databricks and Azure AI Search + RAG — with built-in model evaluation, versioning, and production monitoring.
Image Classification & Detection using Transformers
Fine-tuned Vision and Detection Transformers for real-time image classification and object detection, deployed on an NVIDIA Jetson Nano for on-device live video analysis with OpenCV.
Vehicle Health Monitoring
ML models for vehicle fault classification and remaining-useful-life prediction, using sensor-data preprocessing and anomaly detection to catch issues before failure.
Profile Analytica
An AI-driven document ingestion pipeline combining OCR and NLP for automated resume data extraction, validation, and LLM/BERT-based parsing, summarization, and semantic matching on a scalable MongoDB backend.
Guardrailed LLMs for Test Automation
Fine-tuned, privacy-preserving LLMs (Mistral, LLAMA3) with guardrail configurations and RAG–BM25 retrieval, enabling local, token-efficient inference for SQL and Python code understanding.
Product Recognition System
A visual product-recognition system using ANN and OpenCV for image similarity, with BERT-based vectorization of product names to recommend and match related products from an uploaded photo.