Each of these moved raw, messy input through an automated pipeline into a structured, deployed output.
P&ID Diagram Analyzer
COE-SEA, IIT Kharagpur · Jan–Apr 2025
A pipeline that converts scanned engineering diagrams into structured relational graphs for
industrial data analytics, deployed behind a containerized API for symbol classification at scale.
Input500+ raw P&ID diagram scans (unstructured images)
→
ProcessPyTorch Geometric GNN for symbol classification & graph construction
→
OutputStructured relational graphs, queryable for analytics
PythonPyTorch GeometricFlask
DockerGitHub Actions CI/CD
Automated CI/CD took this from notebook experiment to a production-deployable classification API.
Chatbot Retrieval Pipeline
xLayer Technologies · Aug 2025–Present
An ingestion pipeline that scrapes source content, embeds it, and serves it back through a
production microservice for semantic retrieval.
Input200+ scraped pages of unstructured web content
→
ProcessHugging Face embedding models → Milvus vector store (250+ embeddings)
→
OutputSemantic search / context retrieval for a live chatbot API
PythonFastAPIMilvus
Hugging FaceDockerAWS S3JWT
Runs as a standing microservice, not a one-off batch job — ongoing ingestion, not a single import.
KPI Reporting & Anomaly Detection
Elitech Earth Science · Jun–Jul 2024
Turned raw equipment and geological sensor readings into monitored dashboards and predictive
alerts for an operations team.
InputRaw equipment & geological sensor readings
→
ProcessEDA, anomaly detection, predictive modeling
→
OutputPower BI KPI dashboards + Flask-served predictions
PythonFlaskPower BI
Real-time KPI monitoring cut equipment downtime by 20%; predictive models cut error rates by 12% versus prior statistical methods.
Real-Time Donor Alert System
TechAThon Hackathon · Aug 2025
Event-driven matching: incoming blood requests trigger real-time proximity lookups against a live
donor dataset and dispatch alerts across channels.
Input1,000+ live donor geolocation entries + incoming requests
→
ProcessProximity matching, event-triggered dispatch
→
OutputMulti-channel alerts (Twilio, Firebase, email)
Node.jsExpressNext.js
LeafletTwilioFirebase
Built end-to-end, including the real-time, event-driven matching layer, in a 48-hour build window.
Blockchain Compliance Ledger
IIT Kharagpur Research · Ongoing
Research prototype: mine-site sensor readings are checked against compliance thresholds
on-chain, and only compliant readings mint a certificate — with a backend that turns
those on-chain events back into structured, mapped data.
InputCO2 / PM2.5 / SO2 / noise sensor readings
→
ProcessSolidity threshold check → certificate mint (Hardhat, testnets)
→
Outputethers.js listener → enriched GeoJSON via Express API
SolidityHardhatethers.js
Node.jsExpressPython (RSA signing)
Research/prototype-stage, deployed to Polygon Amoy & Ethereum Sepolia testnets — not a production system.
Stylized low-poly reconstruction of the project's sensor breadboard rig (Arduino, gas sensors, LCD readout) — simplified for the web, not to exact scale. Drag to rotate.