Cross-Industry Privacy-Preserving Predictive Maintenance  

• Preprocessed the SCANIA-X dataset and engineered a hybrid architecture that combines numerical and categorical features of the dataset into transformer embeddings and a TabTransformer for least information loss. Also implemented various differential privacy (DP) algorithms (Spectral-DP, DP-SGD) to protect training data. Designed a global model architecture that would perform well for other PdM datasets that involve numerical, categorical, or any multimodal features in their datasets.
• Infused a Federated training approach based on the flwr library, allowing global model aggregation across heterogeneous clients having different computing power, parameters, and hyperparameters; mirroring real-life industrial equipment.
• Presently it is the best RUL prediction model on the dataset with an MSE of 2725 whilst keeping training data privacy. Developed an advanced MIA that considers white-box, gray-box, and black-box features, along with time-series specific seasonality/trend features; achieved a MIA success rate (AUC) of 49.12% and MIA accuracy of 49.59%, thus solidifying the claim.
• Implemented the DP mechanisms from scratch (including a Renyi Differential Privacy / Moments accountant) and a two-stage model pairing a Transformer time-series encoder with a TabTransformer head. This work is under review at IEEE T-ITS.

Context-Aware Multimodal Knowledge Retrieval  

• Built a context-aware parser to automatically extract and separately process multimodal content (images, tables, equations, graphs, text) and summarize each modality into vector embeddings. Employed modality-specific pipelines and selected appropriate LLMs per content type to improve summary quality; stored embeddings and summaries in ChromaDB using Hugging Face embeddings.
• Designed a multi-vector retrieval strategy that links document summaries to original content for improved context preservation. On query, the system retrieves relevant multimodal content and generates comprehensive, source-cited answers that reference text, table data, and image insights.
• Runs three parallel retrieval pipelines (LLM-summary embeddings, raw-atomic content, and CLIP text-to-image search), parsing documents with Docling and using Groq LLaMA 3.3 70B alongside Gemini 2.5 Flash Vision.

SENTRAL - Multi-Spectrum Stock Analysis  

• Conducted a fundamental analysis of target companies using 14 metrics (P/E, P/B, Debt, ROE, etc.) plus Piotroski F-Score and Altman Z-Score via custom functions. Executed technical analysis using 18 indicators (EMA20/50, SMA20/50, RSI14, MACD, ATR, etc.) and performed peer comparisons across sector companies.
• Collected news via seven APIs and web scraping, filtered items by relevance score, then fed the curated corpus to 10 LLMs (including finance-tuned and SOTA reasoning models) for sentiment extraction and signal generation. Visualized correlations and feature importance to support decision making.
• Applied Transformer and LSTM models to forecast price trends for reference and ensembled sentiment with fundamental and technical indicators to compute buy/hold/sell probabilities. Boosted newbie portfolio worth by +45% over five months.
• Packaged as two production-ready Streamlit apps (a deep single-stock analyzer and a screener.in-style screener) with a 20-strategy backtesting engine, Monte Carlo simulation and automated HTML/PDF report generation.

CodeGen - Autonomous Competitive Programming Solver  

• End-to-end autonomous solver (FastAPI + Google Gemini) built for the Meta Hacker Cup 2025 AI Track; the system behind my Global Rank 10 finish.
• Complete multimodal pipeline that fetches and decodes problem diagrams, generates a Python solution, and validates it against the sample I/O.
• Iterative self-repair feeds failures back to the model and regenerates up to four times; solved upper-medium-to-hard problems (~2200-2500 rating).
• Ships a reusable dev-kit with the prototype history and CP_GEN client tooling.