Projects That Delivered Results
Each one is live, real, and built with a clear purpose — not portfolio filler.
Intelligent Systems
LLMs, RAG pipelines, computer vision, and predictive models
RIS — Recruitment Intelligent System
What it does: End-to-end AI candidate sourcing pipeline — LLM criteria extraction from job descriptions, vector search via Qdrant, LinkedIn scraping, AI-driven scoring with SHAP explainability. 9-stage automation pipeline.
Optimization: Reduced token usage from 0.2M → 5K per iteration (~99% cut) by eliminating LLM-inside-loops anti-pattern and switching to batched inputs with targeted retrieval. Cut run time from 6 min → 90 sec via async orchestration.
Failure & Recovery: When RIS crashed under heavy concurrent scraping load, diagnosed the bottleneck and resolved it — added load balancers, concurrent threading, and mandatory structured logging. System now handles concurrent users reliably.
ResuGenie — Production AI Resume Intelligence Platform
Platform: AWS ECS Fargate + OpenSearch for vector search + Amazon Bedrock LLM inference. Resumes chunked at sentence-level with overlap, embedded and indexed — JD query triggers vector retrieval and Bedrock synthesizes match explanation.
Pipeline: Upload → chunk → embed → index → query → retrieve → synthesize → score. Deployed on AWS until free tier limit; architecture fully production-grade.
JARVIS 2.0 — Autonomous Multi-Agent AI System
Evolution: Started as a Gemini API wrapper (JARVIS 1.0) — in one year, rebuilt it into a full autonomous multi-agent system with a supervisor orchestrator, specialised sub-agents, tool use, and structured reasoning.
Architecture: Supervisor agent delegates tasks to domain-specific workers (research, code execution, web search, file management). Replaced fragile if-else chains with dynamic agent routing and retry logic.
Key shift: LLM outside loops, batched inputs, structured decision trees — same principles applied at Idyllic Services.
Smart House Price Predictor
XGBoost model with SHAP explainability. R²=0.88 on held-out data. 18 engineered features. Users see exactly which features drive their property estimate — chose XGBoost over neural networks for tabular data performance and native SHAP compatibility.
Smart Finance LSTM
Deep learning for financial forecasting. LSTM-based time-series analysis. TensorFlow/Keras. Predictive modeling for financial data. 15+ GitHub stars.
JARVIS 1.0 — My First AI App (Origin Story)
The project that started it all. A Gemini API wrapper with a web interface — simple, but it was my first real AI build. Within a year, this evolved into JARVIS 2.0: a full multi-agent autonomous system with supervisor orchestration, tool use, and structured reasoning. This is where the journey began.
AI Caption Generator
Automatic image captioning using transformers and CNNs. Accessibility support, content management, image indexing. Production-ready Python library. 15+ stars.
Object Detection System
Real-time computer vision. YOLO/Faster R-CNN models. Webcam streams or image processing. High accuracy, low latency. 30+ FPS on GPU. Production-ready.
AI Interview Platform
AI-powered interview simulation. Real interview scenarios. AI feedback and scoring. Comprehensive preparation for technical interviews.
Full-Stack Projects
Backend APIs, real-time features, and production deployments
Raghavendra Swami Mutt — Automated Booking System
Devotees used to stand in queue for 2–3 hours to book a seva. The website reduced this to a 5-minute online booking. Flask + MySQL backend with ACID transactions to prevent race conditions on concurrent bookings. Admin dashboard for real-time management. 90% manual staff effort cut, 95% reduction in per-booking processing time. Real users, 24/7 uptime.
CodeSync — Collaborative Code Editor
WebSocket-based real-time editor. Multiple users code simultaneously with sub-100ms sync. Flask-SocketIO backend with room-based sessions. Operational transformation for concurrent edit conflict resolution. 17+ GitHub stars.
Libraries & Extensions
Python packages and VS Code tools
integration-smoke-test
Lightweight Python library for API health checks. Fast, single-probe testing. CI/CD ready. Structured results, error categorization, suggested actions. MIT License. Live on PyPI.
AI-Powered VS Code Extension
Natural language terminal commands. Execute shell commands by describing what you want. AI-powered command generation. Built from scratch. Productivity multiplier.
Browser Automation & Scraping
Selenium, web scrapers, productivity tools
AutoJobFormFiller
Browser automation for job applications. Selenium-powered intelligence. Fills forms automatically. Resume parsing. Works on LinkedIn, Indeed, Glassdoor. 10-20x speed boost.
eCourts Scraper
Indian eCourts data extraction. Case information, judgments, legal documents. BeautifulSoup + Selenium. Batch processing. CSV/JSON export. Rate limiting. Ethical practices.