Featured Work
A curated collection of projects I've built. Each represents a unique challenge and learning experience.
Synergistic Underwater Image Enhancement
An implementation of fusion-based graph neural networks for underwater image enhancement, benchmarked against traditional methods and CNN baselines on UIEB.
Impact: Improved ResNet-34 U-Net reached 20.75 dB PSNR and 0.8966 SSIM on 90 paired test images in a 10-epoch experiment.
Hotel Booking Cancellation Intelligence
A booking-time cancellation risk model and interactive BI dashboard that recommends deposits or reconfirmations for high-risk hotel bookings.
Impact: Achieved 0.851 ROC-AUC on 23,842 later arrivals using a time-based split and only booking-time features.
Recova — AI Revenue Recovery
An AI agent that detects failed payments, diagnoses revenue risk, and runs bounded recovery workflows using probabilistic judgments and deterministic action gates.
Impact: Demonstrated a Razorpay Test Mode payment capture with policy-gated actions and an auditable recovery trail.
RecallHQ — Team Chat RAG
A RAG system for Slack-style team chat with hybrid vector and keyword search, decision-model reranking, and answers linked to source messages.
Impact: Enforces workspace and channel access controls and withholds answers when citations fail validation.
DepthWizard — Single-View 3D Mapping
Turns a single aerial or satellite image into estimated heights and an interactive 3D scene, with measurement tools and geospatial exports.
Impact: Reached 2.590 m RMSE on 859 GAMUS validation tiles using D4 augmentation and 1.5× input scaling.
Geospatial Flood Segmentation
A 3-class flood segmentation pipeline built from multi-sensor satellite imagery and a fine-tuned IBM/NASA Prithvi EO foundation model.
Impact: Top 40 of 1,000+ teams at ANRF AISEHack 2026.
Crop Type, Health & Yield Estimation
Two-stage optical–SAR fusion pipeline combining Sentinel-2 time series with Capella X-band SAR for crop classification and health estimation.
Impact: Leakage-safe modeling with phenology, spectral, backscatter, and GLCM features.
Traffic Demand Forecasting
An ensemble forecasting workflow for Flipkart Gridlock 2.0 using tuned gradient-boosting models and temporal features.
Impact: 94.56% out-of-fold R² with leakage-safe K-Fold target encoding and lag-96 features.
Sutra — URL Shortener
A full-stack URL shortener with authentication, caching, and a PostgreSQL-backed service layer.
Impact: Sub-500ms redirect latency on a resource-constrained deployment.
Automated Traffic Violation Detection
An in-progress computer vision system for vehicle detection, license-plate recognition, and violation review.
Impact: Designed for low-latency processing and 90%+ OCR accuracy across lighting conditions.