👋 Hello, I'm Shubhro Dev an AI/ML researcher working on large language models, multimodal deep learning, natural language processing, agents and time-series analysis.
🎓 B.Tech in Computer Science and Engineering from RGIPT (an Institute of National Importance, along the lines of the IITs), 2022–2026
🔬 Most recently a Technical Research Intern with the Global DAI R (SDE-IN) team at Siemens, researching the validation and verification of LLM outputs
📝 I like building projects which have real-life impact or which sounds fun to me!!
🌐 Everything in one place: spearb0lt.github.io
All projects, grouped by domain
🧠 NLP, RAG and document AI
| Project | What it is |
|---|---|
| BidRAG | Open-source RAG extraction for RFQs, tenders and contracts, with every answer traceable back to its source page. An open reimplementation of Bidify, the internal Siemens application I worked on: same feature set, lighter, and swappable at every layer (LLM, embeddings, extraction, storage, auth) with a free offline default everywhere. |
| Context-Aware Multimodal Retrieval | A multimodal RAG pipeline over PDFs that retrieves across text, tables, figures and formulas with cited answers. Three retrieval paths run in parallel (summary embeddings, raw atomic content, CLIP text-to-image), sharpened by HyDE expansion and a cross-encoder reranker, with the model always shown the original element rather than its summary. |
| Legal-RAG (Nyaya) | Citation-grounded research over 19,144 passages of Indian statutes and judgments. Every citation is checked against the passage it names before the answer renders, and one that cannot be verified is flagged rather than shown. Hybrid dense and BM25 retrieval answering in about 20 ms, with the whole corpus fitted into 512 MB. |
| Complete RAG Evaluation | A link-verified survey of RAG evaluation covering 547 benchmarks, frameworks, metrics, judge models and papers, regenerated from a single CSV by a verify, enrich, audit and build pipeline. |
🔐 Privacy-preserving ML and time-series
| Project | What it is |
|---|---|
| Privacy-Preserving Predictive Maintenance | Federated, differentially private RUL prediction on the IDA 2024 SCANIA-X challenge. Spectral-DP and DP-SGD written from scratch with a Rényi/Moments accountant, a Transformer time-series encoder paired with a TabTransformer head, aggregated across heterogeneous clients with flwr. Best RUL model on the dataset at MSE 2725, while an advanced membership inference attack gets only 49.12% AUC, no better than a coin flip. |
🤖 Agentic AI and LLM systems
| Project | What it is |
|---|---|
| CodeGen: Autonomous CP Solver | The end-to-end AI solver I built for the Meta Hacker Cup 2025 AI Track, where it placed Global Rank 10. FastAPI and Gemini, multimodal parsing of statements that embed diagrams, and a generate-test-repair loop that validates every candidate against the samples before returning it. Solved problems in the 2200 to 2500 rating range. (dev history and client tooling) |
| LitigatAI | An AI workbench for Indian litigation practice: ten tools covering a matter from the first brief to the final filing, in one FastAPI app with no build step. Bring your own key across five providers, with the credential scoped to a single request and cached under a hash of itself rather than under the provider name. Live app |
| Consumer Commission Complaint Agent | An eight-step intake that turns a plain-language grievance into a file-ready Indian Consumer Commission petition. The jurisdiction, limitation and fee engine never touches a model, so the parts a user would actually rely on are deterministic and identical on every run. Live demo available. |
👁️ Computer vision, accessibility and HCI
| Project | What it is |
|---|---|
| UHCI / SignBridge | A fully offline communication app for Deaf and non-speaking users. Twelve input channels, including morse keyed by a blink or a tap, sign language over a 2,000-word vocabulary, lip reading, eye gaze, air-writing and braille held up to the camera, all feeding one message box. No API keys, no accounts, no network at runtime. Roughly 129 modules, 38.6k lines, 1,044 passing tests, and every performance number measured on real hardware. |
| Face Liveliness Detection | A lightweight, fast anti-spoofing model that tells real faces from fake images and videos. |
| Gesture Video Controller | Webcam hand-tracking to control media playback with single-hand and dual-hand gestures. |
🩺 Biomedical and medical imaging
| Project | What it is |
|---|---|
| Lung Cancer Detection with Soft Attention | Soft-attention deep transfer learning for lung cancer identification from CT scans and histopathology. Published at IEEE ISACC 2025. |
| Federated Cervical Cancer Detection | Cervical cancer detection on Pap smear images, comparing three federated learning strategies. |
| Federated Heart Disease Prediction | Predicting heart disease from clinical biomarkers under federated learning. |
📈 Data analytics and finance
| Project | What it is |
|---|---|
| Olist Marketplace Intelligence | An end-to-end analyst project on Brazilian e-commerce data. Late delivery causes a 1.70-point drop in review score (95% CI [−1.76, −1.67]), survives four robustness attacks including an E-value of 11.17, and the failing leg is carrier transit, not the seller. It closes with a governed-analytics eval where the semantic layer never fabricated an answer and text-to-SQL did so 16.7% of the time. Tested dbt/DuckDB warehouse, about 90 seconds from a clean clone. Live Tableau dashboard |
| TimeWeave | A team hackathon project for American Express (I wrote the bulk of it): cross-channel journey stitching into one identity, churn scoring with per-driver attribution computed in raw feature space, and counterfactual cohorts. Every recommendation is labelled as an observed retention rate, never as causal uplift. (team repo) |
| SENTRAL | Multi-spectrum equity analysis: fundamentals (DCF, Piotroski F-Score, Altman Z-Score), 35 technical indicators, news sentiment from 10 models over a corpus gathered from 13 sources, and LSTM/Transformer forecasting, ensembled into buy/hold/sell probabilities. Nothing ships unbacktested: 20 strategies are replayed under stated assumptions first. |
🧩 Optimisation and operations research
| Project | What it is |
|---|---|
| Non-conflicting Timetable Generator | University timetabling as a constraint satisfaction problem in OR-Tools CP-SAT, with nine lexicographic objective tiers, so "why was my morning preference broken?" is mechanically answerable. Verified by an independent checker that shares no code with the solver, and it diagnoses ("you are 108 teaching hours short") instead of saying infeasible. |
♟️ Open source and packages
| Project | What it is |
|---|---|
| Lichess Essentials | Five chess tools that fix the things I kept running into as a long-time Lichess user: a game analyser, study-to-PDF, a repertoire builder, an opponent scout, and a weakness report that ranks your bad positions by what they cost you beyond your own average. Shipped as six PyPI packages through Trusted Publishing, and running live on one free Oracle Cloud ARM box. (PyPI) |
🧰 Useful tools
| Project | What it is |
|---|---|
| Job-Aligner | Tailors a LaTeX resume to any job description and compiles a submission-ready PDF, with a Chrome extension front end. |
| Text-to-Handwriting Converter | Converts typed text (including tables) into realistic handwriting and exports it as an A4 PDF. |
