I build the unglamorous parts of software that keep the glamorous parts alive: queues that don't lose jobs, retries that don't double-charge, and models whose accuracy survives contact with real data. Final-year CS at NIT Surat, founding engineer intern at Career Passport last summer, and occasionally found arguing with Codeforces about what "easy" means.
Open to: 6-month SDE / backend / AI engineering internships from January 2027, remote or on-site · resume · portfolio
| Languages | Python, TypeScript, JavaScript, Java, SQL, C++ |
| Backend | Node.js, NestJS, FastAPI, Django, Spring Boot, REST APIs, Server-Sent Events |
| Queues and data | PostgreSQL, MongoDB, Redis, BullMQ, Redis Streams, Prisma, Qdrant, Meilisearch, Firestore, MinIO |
| ML | PyTorch, LightGBM, XGBoost, scikit-learn, Hugging Face Transformers, LoRA fine-tuning, ONNX Runtime |
| LLM systems | RAG, retrieval evaluation, rerankers, embeddings, Langfuse, LangChain, Ollama, Groq and Gemini APIs |
| Frontend | Next.js, React, Tailwind CSS |
| Infra | Docker, Linux, Git, GitHub Actions, AWS, GCP, SLURM on A100/H100 clusters |
- Winner, Adani Powermind Hackathon 2026 (team): energy-demand forecasting
- Rank 94 of 19,900+, Amazon ML Challenge 2026: entity resolution over 10M records
- Top 15 (solo), EEG/EMG Foundation Challenge, NeurIPS 2026 workshop
- Top 10 (solo), WWT Unravel 2026
- 1st runner-up, WebOsmotic AI/ML Hackathon 2025
- Top 100, ZS Tech Challenge · finalist at 12 national hackathons
- Chairperson, Cutting Edge Visionaries (Fintech Club), SVNIT · Developer, ACM SVNIT
Codeforces: 1415, Specialist (max 1415)
CodeChef: 1681, 3★ (max 1681)
LeetCode: 1550, 522 solved
- Foundation models for brain signals: decoding EEG for people a model has never seen, from Riemannian geometry to LoRA fine-tuning (my NeurIPS 2026 challenge entry)
- Evaluating retrieval honestly: benchmarking every stage of a RAG pipeline on held-out questions (keyword fusion lost, so it shipped off)
- Distributed job processing: at-least-once delivery, idempotent consumers, and what actually happens when a worker dies mid-job
- Java and Spring Boot, by building something with them rather than reading about them
| What | Why it's interesting | |
|---|---|---|
| Distributed RAG | Self-hosted RAG over PDFs, scans, audio and video, split into 5 services | Retrieval benchmarked stage by stage (nDCG@10 0.706 to 0.724 on held-out questions), duplicate chunks cut from 17% of top results to 0%, and a queue where a crashed worker can't lose a document |
| FinBridge | Offline-first lead capture and call review for field sales (team project) | INT8 sentiment model measured 2.74x faster on CPU with identical labels, and a check for when someone's tone contradicts their words |
| Business Entity Resolution | Amazon ML Challenge 2026, rank 94 of 19,900+ | Matching 10M messy records without comparing 6.7 trillion pairs, and a decision rule built around the metric |
| Jobs180 | Spring Boot job tracker | Ten concurrent retries of the same request create exactly one record, and there's a test that proves it |
| AssetFlow | Asset management (team hackathon project) | Role checks on every request, Ctrl+K search, and real-time notifications that replay what you missed |
| Prometheus | Adani Powermind Hackathon winner (team) | I built the forecaster, then found the leak in my own validation and published the honest number |



