Final-year student at IIT Kharagpur (Integrated Dual Degree, Chemical Engineering and micro-specialization in Artificial intelligence and Application & High performance computing, graduating April 2027). I build production ML systems and do speech/NLP research — the kind that ships with monitoring, gets measured, and defends its numbers. Research → Wav2Vec2 + WavLM-BiLSTM fusion for native language identification from L2 English speech. 0.88 val accuracy across 5 Indian languages. Competitions → Top 0.5% globally, Amazon ML Challenge 2025 (50,000+ teams). 1st place, GC OpenSoft '25 (Deloitte-sponsored). 1st place, GC Data Anytics'25 (Evva health). Math/Algorithms → Codeforces Expert (peak 1612 · _Badassium). AIR 135, Integral Cup 2026 S1 across Probability and Statistics, Linear Algebra & Integral Analysis, Optimization Techniques.
DocuReason RAG: A Grounded Framework for Multimodal Document Retrieval and Reasoning · PyPI
Sponsered Research & Industrial Consultancy (SRIC), IIT Kharagpur · May – Jul 2026
- Architected DocuReason, an enterprise-grade tri-path multimodal RAG framework for grounded document retrieval and reasoning across text, complex financial/scientific tables, and visual PDF documents under the guidance of Prof. Shyamal Kumar Das Mandal.
- Integrated ColPali (colpali-engine) with Hugging Face Transformers (transformers, peft, accelerate) for vision-language document embedding, fused with BM25s hybrid lexical search, Qdrant vector database, and DuckDB tabular execution via Docling & EasyOCR parser pipelines—slashing retrieval hallucination rate by 45%.
- Published and deployed the core framework to PyPI as docureason-framework with async FastAPI inference microservices, Pydantic schemas, and structured benchmarks—authored comprehensive research report.
Advanced Technology Development Centre (ATDC), IIT Kharagpur · May – Jul 2025
Supervisor: Prof. Shyamal Kumar Das Mandal
Benchmarked 7 deep learning architectures on the NISP dataset — from CNN and ResNet-18/50 baselines to transformer-based SSL fusion models — across 5 Indian L1 languages (Hindi, Telugu, Tamil, Kannada, Malayalam).
Final model: Frozen Wav2Vec2 + WavLM encoder layers → BiLSTM head. 0.88 validation accuracy, state-of-the-art on this task. UMAP embeddings revealed tight, well-separated Telugu/Malayalam clusters and quantifiable Hindi-Tamil phonetic overlap — interpretable signal beyond the classification score.
| Project | Result | Stack |
|---|---|---|
| DocuReason RAG Framework | DocuReason RAG: A Grounded Framework for Multimodal Document Retrieval and Reasoning | ColPali, PyTorch, Transformers, FastAPI, Qdrant, DuckDB, BM25s, Docling, PyPI |
| Multimodal Price Prediction | Top 0.5% Amazon ML 2025 · SMAPE 25.45 · 469ms p95 latency · SHA-256 model fingerprinting · 46+ MLflow runs |
PyTorch · SBERT · CLIP · XGBoost · LightGBM · CatBoost · Ridge · FastAPI · MLflow · DVC · DagsHub |
| RAG Employee Support Platform (repo uploading) | 1st Place GC OpenSoft '25 (Deloitte) · −35% resolution time · −40% hallucination rate · −30% cost/session · +50% API throughput |
LangChain · OpenAI · FAISS · Chroma · FastAPI · Docker · Prometheus · Grafana · OpenTelemetry |
| Text Summarization System | Fine-tuned BART pipeline · end-to-end CI/CD deployment | PyTorch · HuggingFace · FastAPI · Docker · GitHub Actions |
| Portfolio Website | Role-personalized SPA · FastAPI backend · live ML data context | React · TypeScript · Vite · Tailwind · FastAPI |
| Technology | Name |
|---|---|
| NLP / Speech AI |         |
| ML / Deep Learning |         |
| MLOps & Infrastructure |         |
| Backend / Serving |     |
| Languages | |
| Cloud & Tools |