About
Fourth-year CSE (Data Science) student at VIT Chennai building practical AI/ML systems for real-world applications. Experience spans multimodal document extraction, explainable machine learning, and deployed applications — including a co-authored invoice extraction system (manuscript under consideration) and Edge Defense, a deployed explainable intrusion detection system. Currently building VolatilityIQ, an explainable financial volatility forecasting system combining GARCH models, deep learning, and SHAP-based regime analysis on Indian equities.
Skills & Expertise (27)
Work Experience
RECIQ - Multimodal Receipt Extraction System
VIT Chennai
Sep 2025 - Present
Achieved 83.04% Active Exact Match and 99.0% parse rate on 200 held-out real-world Tanglish invoice images, outperforming the strongest baseline by +74.9 percentage points - Manuscript under consideration. Implemented tokenizer surgery injecting 174 schema field names as atomic vocabulary tokens, eliminating subword fragmentation and reducing final training loss by 274× compared with vanilla LoRA fine-tuning. Applied LoRA (r=16, α=32) across all 28 decoder layers of Qwen2-VL-7B on 7,124 synthetically generated invoices across 9 vendor domains; converged in 1,299 steps to a final loss of 0.0126 on an H100 GPU. Designed a grammar-constrained beam decoder guaranteeing syntactically valid JSON at every generation step, with a downstream GNN anomaly detector achieving macro F1 of 70.0% for financial auditing.
VolatilityIQ – Explainable Volatility Forecasting for Indian Equities
Independent Project
Present - Present
Designing a multi-phase explainable volatility forecasting system for Indian equities, combining classical statistical models (GARCH) with deep learning and SHAP-based explainability to study how model reasoning shifts between calm and crisis market regimes. Incorporating earnings-call transcript signals - structured language features rather than a single sentiment score - to test whether text-based signals improve forecasts around earnings events. Exploring cross-stock volatility spillover using graph neural networks with attention-based explainability to capture how shocks propagate across related stocks. Applying rigorous statistical validation (permutation testing, FDR correction, Diebold-Mariano tests) throughout, so every finding - positive or null - is defensible.
AI Network Intrusion Detection System
Edge Defense
Jan 2026 - Apr 2026
Built and deployed a full-stack network intrusion detection system using XGBoost, achieving 99.57% accuracy and 0.9998 ROC-AUC, with SHAP-based explainability on every prediction for per-alert feature attribution. Deployed end-to-end with a FastAPI backend and a React dashboard featuring live traffic analysis and batch CSV processing — live at edge-defense-ui.vercel.app.
CareConnect — Smart Healthcare Recommendation System
Personal Project
Jan 2025 - Jan 2025
Built a healthcare platform matching patients to providers based on symptoms, location, and specialization using a Django backend with SQLite database and Tailwind CSS frontend. Implemented secure authentication, appointment scheduling, and provider recommendation logic.
Education
B.Tech in Computer Engineering (Data Science Specialization) - Vellore Institute of Technology
2023 - 2027 · Afghanistan
Certifications
CalmWeave: A Tri-Layer Adaptive Smart Vest for COPD Respiratory Support
NCSFI 2026 Proceedings · 2026
Interested in this developer?
Profile Score Breakdown
Profile Overview
Availability Details
Relocation
Open to Relocation
Skills (27)
Click a skill to find developers with the same skill
