MS Computer Science at UW-Madison. I work on LLMs, ML systems, and the engineering underneath them. Most recently a Data Science Intern at Cloudflare. Before graduate school, 2+ years at Oracle building backend services and data infrastructure.
Looking for new grad and early career SDE and ML engineering roles. Available January or June 2027.
Knowledge MCP Server · Python, MCP SDK, LangChain, Vector DB
An MCP server that lets Claude Desktop and Cursor agents pick their own corpus: create, add, list and query tools over stdio. Retrieval fuses dense vector similarity with traversal over an entity-relation graph the LLM extracts from the documents, with entity resolution so traversal follows real relations instead of duplicate nodes. Tuned on nDCG.
Open source: RSEM (merged upstream) · C++17, CMake, Python, CI
Modernized a widely cited RNA-Seq quantification tool from C++98 to C++17 and removed its legacy Boost dependencies, with a drop-in RNG so existing analyses still reproduce bit-for-bit. Added the codebase's first automated test suite across four aligners with cross-platform CI, then ported the Perl drivers to Python 3. PRs: #233 · #234 · #235
CoT-Faithfulness · Python, PyTorch, LLM APIs
A behavioral and mechanistic audit of chain-of-thought faithfulness across Llama 3.1 8B, Gemma 3 12B, and Claude Opus 4.6. Four CoT perturbations re-injected via assistant prefilling, plus attention analysis. Faithfulness falls as capability rises: 0.49 → 0.30 → 0.09.
Certified-Temporal-GNN · PyTorch, Randomized Smoothing
Certified robustness for temporal graph neural networks. Replaces the per-window Bonferroni union bound with one shared edge-flip realization under a single Clopper-Pearson bound: sound under Neyman-Pearson, strictly tighter, and lifts certified accuracy from 26.7% to 68.3% on the real Elliptic Bitcoin graph.
Intelligent-Bistro · TypeScript, React Native, Expo, Express, Zod
A mobile ordering app where guests manage a cart by talking to an assistant. Claude parses natural-language orders into validated tool calls. UI taps and AI calls share one cart-mutation path, so the model's view never drifts from the user's, and pricing is always computed server-side.
SynthShield · PyTorch, LoRA, Opacus, FastAPI, Next.js, Cloud Run
Turns sensitive tabular data into synthetic clones with formal privacy guarantees. T5-small fine-tuned with LoRA under DP-SGD, reaching ε≈4. Built in 24 hours at CheeseHacks 2026.
- Languages & Core: C · C++ · Python · Java · JavaScript · TypeScript · SQL · HTML/CSS · MATLAB
- Backend & Distributed Systems: Spring Boot · Microservices · REST APIs · Node.js · Express.js · gRPC · Kafka · Redis
- ML & Modeling: PyTorch · TensorFlow · Keras · Scikit-learn · Hugging Face · Pandas · NumPy · MLflow · Braintrust
- LLM & Agents: RAG · MCP · AI Agents · Vector Search · LangChain · Fine-Tuning (LoRA) · Claude Code · Cursor · Codex
- Data, Cloud & MLOps: Spark · BigQuery · Elasticsearch · Oracle · MongoDB · AWS · GCP · OCI · Docker · Kubernetes · Jenkins (CI/CD) · Git · Linux
Experience
- Cloudflare, Data Science Intern · Jun 2026 - Aug 2026
- UW-Madison Data Science Institute, Open Source Developer · Feb 2026 - Jun 2026
- Oracle, Applications Developer · Jun 2023 - Aug 2025
Education
- UW-Madison, MS Computer Science
- IIT Kharagpur, B.Tech (Hons.) Mechanical Engineering, Minor in Mathematics & Computing · 2019 - 2023
Publication
- DARSI: A deep auto-regressive time series inference architecture for forecasting of aerodynamic parameters, Journal of Computational Science (Elsevier), 2024
- Email: jmahajan2@wisc.edu
- LinkedIn: linkedin.com/in/Jeevesh28