Agentic AI OS for Supply Chain Intelligence

An LLM-orchestrated multi-agent platform that autonomously manages global supply chains — demand forecasting, supplier risk scoring, inventory rebalancing, logistics optimization, and sustainability tracking, all coordinated through a hierarchy of 8 specialized AI agents.

What makes it different:

• Orchestrator dynamically routes natural language queries to the right agents
• Supplier risk scoring runs through a trained scikit-learn ML pipeline — fully bypassing the LLM for deterministic, explainable predictions
• Smart mock-fallback: all 12 LangChain tools execute on real data even when Ollama is offline
• Real-time agent status polling every 3s; inter-agent message bus with ring-buffer logging

8 Agents · 12 Tools · 10 Workflows · MongoDB · FastAPI · React 19

FastAPI LangChain LangGraph Ollama (llama3.2) MongoDB React 19 Tailwind scikit-learn

Responsible AI Compliance Platform via RAG

Upload an organizational policy or document → select regulatory frameworks → receive structured compliance assessments with per-article risk levels, metric scores, and actionable recommendations — grounded in actual regulatory text via RAG.

What makes it different:

• FAISS + LangChain LCEL retrieves exact regulatory passages per framework (DPDP · GDPR · EU AI Act · NIST AI RMF)
• Analysis runs asynchronously in the background; users fill the governance questionnaire in parallel
• Scores computed manually from per-article results — eliminates LLM arithmetic hallucinations
• Full SQLite-backed audit trail with 10+ distinct event types at every pipeline stage

Dual LLM · FAISS · Async Analysis · Governance Module · Full Audit Trail

FastAPI LangChain (LCEL) FAISS Ollama / Gemini 1.5 React 18 SQLite nomic-embed-text

Enterprise AI @ Atos (Eviden) — Production Project

Contributed to Atos' first GenAI-powered enterprise modernization initiative. Led the LLM API migration of 39 AWS Lambda functions from Claude Sonnet 3.5 → 4.0, performed dependency refactoring (langchain-community → langchain-aws), implemented regex-based Markdown post-processing, and designed DevOps artifact generation (Dockerfiles, Helm Charts).

FastAPI AWS Lambda AWS Bedrock LangChain boto3 GitHub Actions EC2

Semantic Space Knowledge Retrieval

LLM-powered conversational bot using RAG architecture for natural language querying of space research data. Domain-specific chunking, hybrid retrieval, and LangSmith observability.

Flask LangChain LangGraph ChromaDB LLaMA 3.2 (1B) nomic-embed-text LangSmith

Voice-Based Fraud Prevention System

ML + voice feature extraction for real-time caller verification and fraud detection.

📄 IEEE CONIT 2024, Bangalore DOI: 10.1109/CONIT61985.2024.10626085

Django Machine Learning Voice Analysis Feature Extraction

Lakehouse Engineering on Databricks

Medallion architecture (Bronze → Silver → Gold) over 38M+ NYC taxi trip records. KPI tables, window functions, demand forecasting insights.

Apache Spark Databricks PySpark Delta Lake Spark SQL

Deep Learning Structure Detection

YOLOv8 trained on the DocLayNet dataset for real-time document structure classification via Flask REST API.

Accuracy: 0.94 mAP@0.5

YOLOv8 PyTorch OpenCV Flask DocLayNet

AI Review Summarization Platform

End-to-end pipeline for sentiment scoring + AI-generated summaries of customer reviews. RoBERTa polarity scoring + Gemini summarization.

Django RoBERTa Gemini 1.5 Flash Pandas Matplotlib

Real-Time Athletic Intelligence

Sensor-driven exertion monitoring using NodeMCU ESP8266 + ML regression. Random Forest Regressor achieving R² = 0.997 on live sensor data.

NodeMCU ESP8266 ThingSpeak Random Forest Python Matplotlib