My Projects
Building production ML systems that solve real problems. Focus on transparency, explainability, and honest evaluation of limitations.
Featured Projects
PROD-iQ: AI System for Startup Financial Analysis
Reducing LLM Hallucination with Delta State Graph Architecture
The Problem
In 2025, over 11,300 startups shut down in India. Most founders lack quantitative tools to validate financial assumptions. Standard LLMs confidently generate revenue forecasts even when input data is incomplete, leading to confident-but-wrong predictions.
The Solution
PROD-IQ uses a Compound AI Architecture where the LLM orchestrates but never predicts. The system delegates math to specialized ML models and validates against databases. Result: 82% reduction in hallucination instances.
Core Innovation: Delta State Graph
A state-tracking layer that distinguishes KNOWN (user-provided), MISSING (gaps),
ASSUMED (benchmarks), and INFERRED (predicted) data.
Prevents the LLM from fabricating numbers to fill gaps.
Technical Highlights
- Fine-tuned LLaMA-3.2B with 4-bit quantization (custom orchestration layer)
- 5 specialized ML models: Success classifier, revenue estimator, survival predictor, break-even calculator, traction timer
- Trained on 42,500 real startup samples (web-scraped, not Kaggle)
- 387 engineered features (financial, operational, engagement metrics)
- Hybrid architecture: Vector DB (ChromaDB) + SQL (MySQL) + LLM + MCP server
- Cross-validation via vector similarity + SQL benchmarks
- Honest evaluation: 65% reliability with explicit limitations
Tech Stack
TITAN: Planetary Supply Chain Intelligence
Real-Time Ripple Effect Prediction with Graph Neural Networks
The Problem
In 2021, a fire at a Japanese chip factory caused ₹42 Crore ($5M) losses for downstream manufacturers. Why? They lacked Tier-N visibility—couldn't see dependencies beyond their direct suppliers. By the time they knew, it was too late.
The Solution
TITAN is a real-time supply chain intelligence platform that predicts downstream "ripple effects" of disasters with 94% accuracy. Uses a global logistics knowledge graph (Neo4j) to trace dependencies 6 tiers deep. Provides 14-day early warnings for stockouts.
Core Innovation: DCBA Algorithm
Dynamic Context Budget Allocation - Orchestrates 5 AI models without exceeding 8K token limits. Dynamically allocates context budget: GraphRAG (70%), VectorRAG (20%), CRAG (10%) based on query complexity.
Technical Highlights
- Neo4j graph database: 200,000+ nodes, 7,200+ relationships (factories, suppliers, ports, routes)
- DCBA algorithm: Multi-model orchestration with dynamic context allocation
- 6-tier deep analysis: Multi-hop graph queries trace dependencies across supply chain
- Predicts stockouts 14 days in advance with 94% accuracy
- "God Mode" simulation engine: Stress-test supply networks under disaster scenarios
- Hybrid architecture: GraphRAG + VectorRAG (ChromaDB) + CRAG + Gemini 1.5
Tech Stack
Other Projects
Puddle (LMS)
Django | Tailwind CSS | Gemini AI
Learning management system with dual dashboards for staff and students, featuring course uploads, enrollment tracking, and Gemini AI chatbot.
Chef Claude AI
React Vite | HuggingFace | Cloudinary
Cloud-deployed AI recipe generator that creates markdown recipes from leftover ingredients using HuggingFace Transformers API.
EIDA Gamified Learning
React | TensorFlow | YOLO | MongoDB
AI-powered educational platform for children aged 6–13 with interactive games, real-time object recognition, and voice/text AI chat.
SenseLink Communication
Raspberry Pi | OpenCV | Computer Vision
Assistive wearable using Raspberry Pi and computer vision to detect objects and relay live audio instructions for visually impaired users.