Project Associate · IIT Madras Healthcare Technology Innovation Centre
Building the GPU morphing stack for imageless total knee arthroplasty.
Before the research work I spent a year shipping applied ML — always taken past the notebook and served behind a real API, because a model nobody can call is a model nobody uses. That habit is the throughline. I like the part of the job where a result becomes a thing that runs.
| 🎓 | B.E. Computer Science — VTU Bengaluru · CGPA 9+/10 · Class of 2026 |
| 📍 | Chennai & Bengaluru, India |
| 🔬 | Currently: CUDA · Holoscan SDK · neural signed distance fields |
| 💬 | Ask me about GPU pipelines, point-set registration, or why your test suite should include a determinism check |
| 🌱 | Open to software and ML engineering roles |
Six repositories. 419 tests. Every number below is measured and reproducible by cloning.
🎓 AI Tutor
Ask an LLM to tutor and it answers the question — exactly the wrong behaviour. This one is grounded in a knowledge dependency graph, so it knows precisely what it's withholding. Get something wrong and it works out which of 33 named misconceptions produced that answer, then asks the question built to break it.
TypeScript · Next.js 15 · Bayesian Knowledge Tracing |
🌱 EcoEcho
Most tools are a lookup table keyed on a region name. This retrieves over a pgvector index and folds live weather at your coordinates into the query. Then re-ranks against each species' real climate envelope — because similarity alone happily recommends a tree needing 2000mm to someone with 450mm.
Next.js 15 · Supabase · pgvector · Clerk |
The radio isn't the hard part. A C++ layer collapses 215 raw reads into 8 arrivals; routing knows a staircase from a corridor; and the announcement policy decides when to stay quiet — because a device that narrates constantly gets switched off.
C++17 · Python · accessibility-aware Dijkstra |
Spam classification at 99.5% accuracy (1 false positive in 966). Car prices cross-validated on 201 rows. And a housing model scoring R² 0.81 on a random split — −0.13 on a geographically blocked one. That negative result is the most useful thing in the repo, and it ships in every API response.
Python · scikit-learn · pandas |
22.5% more range on an urban cycle — and modelling why it isn't more is the finding. Peak braking hits 32 kW against a 2.5 kW rating, so half of all braking time is spent above what the machine can absorb.
Python · NumPy · drive-cycle simulation |
🐾 PetCare
A vaccination schedule isn't a list of dates — it's a series of dependent intervals. And counting doses isn't enough: three doses finished at 14 weeks is not protection, because maternal antibodies still block the vaccine. The app says so, in words.
Vanilla JS · no build step · offline-first |
| Share | Language | Where it lives |
|---|---|---|
| 44% | 🐍 Python | Applied ML · EV Recovery · RFID Navigation |
| 34% | 🔷 TypeScript | AI Tutor · EcoEcho |
| 13% | 🟨 JavaScript | PetCare |
| 5% | ⚙️ C++ | RFID reader firmware |
| 4% | 🎨 CSS · SQL · HTML · CMake | Styling · pgvector schema · build |
419 tests across six repositories — 95 · 84 · 80 · 66 · 55 · 39, plus 45 C++ checks.
Measured from 559 KB of source via the GitHub API — not estimated.
🔬 Project Associate — IITM HTIC MedTech Incubator, IIT Madras · Jun 2026 – Present
- worked on the GPU morphing stack behind an imageless total knee arthroplasty navigation system — neural signed distance fields, Gaussian RBF field-space warping, and a Holoscan operator graph that closes the loop inside a single surgical frame.
CUDA Holoscan SDK Neural SDF C++ Python
🧪 Project Intern — IITM HTIC, IIT Madras · Feb 2026 – Jun 2026
- Built the Holoscan SDK operator graph for GPU scheduling, keeping the intraoperative path free of host round-trips
CUDA PyTorch Benchmarking
📊 Data Science Intern — Oasis Infobyte · Mar 2025 – Jul 2025
- Shipped three end-to-end ML projects — spam classification, price regression, housing pipeline — each taken past the notebook and served behind a REST API
- Treated reproducibility as the deliverable: every submission documented, re-runnable, and under version control
scikit-learn Flask FastAPI pandas
| 🥈 | Runner-Up — Paper Presentation ICRCCT 2K24 · temporal graph neural networks for predictive maintenance · ADVAYA trophy |
| 🎖️ | Young Turks 2025 — Certificate of Merit 96.32nd percentile in India's largest campus skill contest |
| 🌟 | Aspire Leaders Program 2025 Selected cohort · Aspire Institute, founded by Harvard Business School faculty |
| 📜 | Oracle Certified Generative AI Professional · UiPath Automation Developer Associate AWS Academy — Cloud Foundations & Cloud Security Foundations |
| 👩💻 | Google Developer Group member · SHEfi Scholar · GSSoC open-source contributor |
