Electrical Engineering undergraduate at IIT Bombay, Class of 2029, with a minor in Machine Intelligence and Data Science. I research continual learning and representation learning, and build projects that connect machine learning with practical problems.
Email · Research code · Projects · Try RareBridge
Three first-author research works with Anirban Pal. Each row links the extended manuscript's code separately from its related workshop study; the six codebases represent versions of three works.
| Research question | Extended manuscript code | Related workshop code |
|---|---|---|
| What can feature alignment recover when correspondences are unknown? | Test-Time Feature Alignment under Uncertain Correspondence · Under review at ICLR 2027 | Beyond Marginals · NewInML, NeurIPS 2026 |
| Why can better feature reconstruction worsen retrieval? | Reconstruction-Based Selection Can Hurt Retrieval Compatibility · Under review at TMLR | The Interface Contract · NeurReps, NeurIPS 2026 |
| How does target assignment affect learning from replay? | How Target Assignment Shapes Experience Replay · Under review at TMLR | Deconstructing Gradient Reuse · NewInML, NeurIPS 2026 |
ICLR: International Conference on Learning Representations. TMLR: Transactions on Machine Learning Research.
The repositories retain experiment implementations, configurations, compact evidence and reproduction instructions. The five newly packaged releases preserve their original artifacts and checksum manifests inside artifact/; the existing Interface Contract artifact remains intact. Historical titles may remain inside frozen releases.
| Project | What it explores | Status and starting point |
|---|---|---|
| RareBridge | Source-grounded AI to help rare-disease patient organisations inspect research resources and prepare collaboration briefs. My work focused on AI integration and brief generation. | Team hackathon prototype · Live demo |
| EnerGenius | Adaptive fan and lighting control combining occupancy, environmental sensing, occupant preferences and reinforcement learning. | Ongoing development · Architecture and setup |
| AgriSure AI | Functional farm/financial assessment API with revocable consent, temporal feature extraction, an explainable prototype index, persistence and audit checks. | Rebuilt local prototype · API setup |
| Inventory Rewired — Stratify | Chronological demand forecasts, a dated replenishment ledger, batch expiry and fair policy comparisons across three stores and ten SKUs. | Rebuilt implementation · Checked case report |
| Techfest Algorithmic Trading | Momentum and inverse-volatility allocation with next-open orders, cash/share accounting, fees, slippage and executable stops. | Rebuilt research backtester · Synthetic demo |
Hindi/Hinglish TTS evaluation and normalisation — Inter-IIT bootcamp work covering baseline audio evaluation, participant listening annotations, handwritten text-normalisation rules and a prepared GPU synthesis workflow. The README distinguishes completed work from pending audio experiments.
These are forks of Climate Change AI teaching materials, with the original authors and attribution retained. They are listed separately from original project work.
| Topic | Tutorial fork | Original material |
|---|---|---|
| Building-energy control | CityLearn | Upstream |
| Geospatial ML | Agricultural monitoring with FTW | Upstream |
| Climate modelling | Downscaling climate projections | Upstream |
| NLP | Climate policy analysis | Upstream |
| Introductory ML | Tsunami alerts from earthquake observations | Upstream |
- RL-based-energy-efficiency: earlier energy-control placeholder; the active implementation is EnerGenius.
- scikit-learn: fork of the upstream machine-learning library, listed for reference rather than as an original project.
Programming: Python, C++ · ML/data: PyTorch, NumPy, pandas, scikit-learn · Hardware description: VHDL, Verilog · Other tools: Git, LaTeX, Excel, Power BI