8+ Shipped Projects  ✦  4 Research Tracks  ✦  AI Systems Lab (Flagship)  ✦  AI Systems + Backend  ✦  VIT Bhopal

Engineering complete AI systems — bridging raw data, evaluated models, robust APIs, and products people can actually use.

         

I'm a Computer Science undergraduate specializing in AI & ML at VIT Bhopal. Most of my energy goes into the practical, often-overlooked half of machine learning: making models actually hold up in production.

Training a model inside a Jupyter notebook is usually just the first 20%. The remaining 80% is where engineering counts — writing type-safe APIs, sanitizing unpredictable user inputs, designing resilient database schemas, and crafting interfaces that people can seamlessly interact with. I enjoy owning the entire lifecycle: taking a raw problem, structuring the dataset, evaluating the model, wrapping it behind a documented service, and shipping it.

• AI Systems Lab Active Flagship
  A working reference implementation of AI infrastructure: an inference gateway with SingleFlight request deduplication, circuit breakers, and hybrid BM25 + dense-vector retrieval fused with Reciprocal Rank Fusion.
• Libris Active Build
  An enterprise-style library management backend on Spring Boot — layered architecture, Google OAuth2/OIDC SSO, dual-theme frontend, and CI-enforced test coverage.
• AthenaeumAI
  A RAG-grounded AI study platform turning PDFs into quizzes and flashcards, with SM-2 spaced repetition driving an adaptive review queue.
• Backend & Systems Design
  Deep-diving into distributed systems principles: REST API contracts, caching strategies, database indexing, and write-heavy data consistency.
• Computer Vision & Multimodal AI
  Lightweight vision and voice models (YOLOv8/11, SAM, OpenCV, Whisper) optimized for real-time inference with graceful degradation.

Languages

Backend & Web

AI / ML & Computer Vision

Data & Infrastructure

A working engineering reference for AI infrastructure — not a notebook, an actual gateway.

An inference gateway applying TTL+LRU caching, SingleFlight in-flight request deduplication, micro-batching, circuit breaking, and provider failover across Anthropic, OpenAI-compatible backends (Ollama, vLLM, DeepSeek, Groq), and a deterministic MockProvider. A hybrid retrieval pipeline fuses BM25 lexical search with dense-vector search via Reciprocal Rank Fusion, backed by a RAG evaluation suite (precision, recall, faithfulness). Ships with an interactive playground, Prometheus metrics, and 40 tests. Benchmarked: 200 concurrent duplicate prompts drop from 297.7ms to 45.5ms wall time via SingleFlight — a 99.5% deduplication rate.
Python FastAPI Redis-style Caching RRF Retrieval Prometheus  • 

Enterprise-style library backend, built the way production systems are built rather than the way tutorials build them.

A layered REST backend (Controller → Service → Repository) on Spring Boot 3.5 with MySQL and Flyway-versioned migrations, paired with a dual-theme (Ember Dark / Verdigris Light) vanilla ES-modules frontend. Session-based auth via BCrypt plus opt-in Google OAuth2/OIDC SSO, strict role-based access across Admin, Librarian, and Student, CSRF protection, structured JSON logging, and interactive Swagger/OpenAPI docs. 20 tests across 4 test classes with JaCoCo enforcing ≥70% coverage, deployable to Railway, Render, Fly.io, or Hugging Face Spaces.
Java 21 Spring Boot MySQL Flyway Docker  • 

RAG-grounded AI study system that turns uploaded PDFs into quizzes, flashcards, and adaptive review.

A chunk-based ingestion and semantic retrieval pipeline feeding Groq's Llama 3.3 70B for generation, with SM-2 spaced repetition and an exponential-decay retention model driving the adaptive review queue. Generation runs through a Redis/BullMQ background-job layer with typed enqueue errors and retry-safe validation, backed by 118 unit and 21 integration tests, Playwright E2E coverage, and a GitHub Actions CI pipeline.
React Node.js MongoDB Redis Groq LLM  • 

End-to-end retail intelligence: computer vision, NLP, and a hybrid chatbot behind one FastAPI gateway.

Face recognition (OpenCV + LFW-trained embeddings, privacy-by-design — stores mathematical encodings, never raw images), MobileNetV2 product classification, TF-IDF + Logistic Regression sentiment analysis on customer reviews, and a hybrid rule-based + ML-intent chatbot, all behind a shared singleton model-loading pipeline. Ships with a React + Vite + TailwindCSS dashboard, live webcam capture, and Chart.js analytics — deployed with CI/CD across Render and Vercel.
FastAPI OpenCV React Docker GitHub Actions  • 

Geospatial ML platform for prioritizing urban civic complaints.

KNN spatial-density features, DBSCAN hotspot clustering, and an XGBoost classifier, cross-validated (5-fold) on a hybrid 1,600-record synthetic and 5,469-record real NYC 311 dataset for 95.7% accuracy. An ablation study showed spatial-density features alone lifted Decision Tree accuracy by 12.9 points, and a zero-shot Bhopal→NYC transfer experiment quantified the cross-city domain shift. Shipped behind a React/Firebase reporting app with Leaflet hotspot mapping.
React Firebase Python XGBoost Scikit-learn  • 

Face-gated voice assistant that triggers only on visual user presence.

OpenCV Haar-cascade presence gating, Whisper voice transcription, and a deterministic 5-category intent router with robust fallback handling.
Python OpenCV Whisper MediaPipe Tkinter  • 

Browser-native voice shopping list — no backend, no login, just the Web Speech API.

A command parser turns natural speech into structured actions (add, remove, update quantity, filter by brand or price), with history-based and seasonal product suggestions persisted in localStorage. Voice recognition supports English, Hindi, Telugu, and Tamil, with a text-input fallback when speech isn't available.
JavaScript Web Speech API Firebase Hosting  • 

Computer vision pipeline for multi-crop pathology identification.

Segment Anything Model (SAM) for leaf/stem isolation coupled with custom-trained YOLOv11 detectors for wheat, tomato, and cotton disease diagnosis.
YOLOv11 SAM TensorFlow Lite  • 

• GA-ET-IVCFS-ANFIS — Neuro-fuzzy product-ranking architecture under high uncertainty, co-authored with faculty (R² = 0.98).
• Civic Complaint Prioritization — Hybrid synthetic + NYC 311 evaluation methodology for geospatial civic triage.
• Multimodal AI Assistant — Solo-authored IEEE-style research on real-time vision-speech co-activation.
• Parkinson's Biomarker Mining — Acoustic feature extraction pipeline with optimized SVM classifier (F1: 0.935, ROC-AUC: 0.945).

 

•   Currently reading: Designing Data-Intensive Applications by Martin Kleppmann
•   Current rabbit hole: Model Context Protocol (MCP) & autonomous coding agent architecture
•   Engineering bias: End-to-end usable systems over isolated model demos

900+ problems solved across five platforms, backed by a public solutions & templates repository.

TCS CodeVita Season 13 — Global Rank 3,633.  •  Oracle Certified Foundations Associate — Agentic AI, Oracle University.  •  NPTEL (IIT Kharagpur) — Cloud Computing (Elite + Silver, top 2%) and Introduction to IoT (Elite + Gold, top 5%).