Led a 6-person team to build an AI-powered forgery detection system for Indian identity documents (Aadhaar, PAN, Driving License), reaching 85–90% detection accuracy across a multi-layer forensic verification pipeline. Co-authored a research paper on the methodology, published in JETIR (Vol. 11, Issue 11, Nov 2024).

• Architected a 3-tier microservice system — React frontend, orchestration API, independent ML service — so the AI pipeline scales separately from the core app
• Built an ensemble verification engine: ResNet-50 layout embeddings, EasyOCR with custom character-correction, ORB copy-move detection, Error Level Analysis, EXIF forensics, and Verhoeff checksum validation
• Designed a weighted scoring engine combining all module outputs into Authentic / Suspicious / Forged classifications
• Containerized the full stack with Docker for reproducible deployment

A production-style backend for managing financial records, with a clean layered architecture (controller–service–repository) for maintainability and scale.

• JWT-based authentication and role-based access control across three permission tiers, securing 15+ REST endpoints
• Dashboard analytics APIs delivering category-wise aggregation, monthly trends, and real-time income/expense/net balance summaries
• Filtering, pagination, and full-text search across records using Spring Data JPA specifications
• Global exception handling, unit tests with JUnit 5 & Mockito, and a documented Postman collection

A backend system to ingest, deduplicate, and process time-series manufacturing events at scale, exposing analytical REST APIs for factory and machine-level insights.

• "Latest-received-wins" deduplication strategy guaranteeing a single canonical record per event
• Thread-safe concurrent processing without in-memory locking — relying on DB constraints, @Transactional atomicity, and batched writes
• Optimized for 1,000+ events/sec via batch APIs, HikariCP pooling, and indexed PK lookups — ~30% efficiency gain
• Handled real-world edge cases: anomalous timestamps, stale duplicates, invalid values

A university management application automating term grant slip generation, cutting manual administrative effort by ~40%. Awarded an official Government of India copyright on the underlying algorithm.

• Applied OOP principles across a layered application design
• Added search and filtering to improve admin efficiency and student experience