- 🎓 Final-year B.Tech CSE (AI & Robotics) at VIT Chennai, graduating 2027 · CGPA 8.66
- 💼 Software Engineer Intern @ AdConvergence Mediatech (May–Jul 2026): traced recurring HTTP 500s across 15+ endpoints to database connection-pool exhaustion and fixed them, then consolidated the API layer into a single RESTful service
- ☁️ AWS Certified Solutions Architect – Associate · Selected for Amazon ML Summer School 2026
- 🧠 LeetCode Knight (top 4.31%, contest rating 1914) with 311+ problems solved
- 🎯 Open to full-time SDE, Cloud, AI and Systems Software roles for 2027
| Project | What it does | Stack | Links |
|---|---|---|---|
| StrataDB | LSM-tree key-value store built from scratch: write-ahead log with group commit, SSTables with bloom filters, leveled compaction and crash-consistent recovery. Benchmarked against LevelDB 1.23 on 10M keys across 13 workloads; perf flame graphs traced up to 60% of CPU to a byte-at-a-time CRC, and hardware CRC raised most in-memory reads 49-146% and random writes 73%. | C++20, CMake, perf, ASan, TSan, UBSan | Code |
| Graywatch | Go daemon that finds hosts passing health checks while failing real traffic, on a simulated 20-container fleet. Flagged all 9 injected fault types in 81/81 trials at a ~3 s median (7 s for DNS faults), where shallow health checks caught 3/27 and deep checks 12/27. Automated HAProxy drains cut p99 latency from ~2,000 ms to ~30 ms under CPU and I/O throttling. | Go, eBPF, HAProxy, Docker, Prometheus, Grafana | Code |
| kernc | Compiler with no LLVM for a C-like GPU kernel language, targeting NVIDIA PTX. Final SASS is within 2 instructions of nvcc on every benchmarked kernel. Runs 205 tests including GPU end-to-end on an RTX 3060; testing found and fixed 16 real compiler bugs. | C++17, PTX, CUDA Driver API, GoogleTest | Code |
| cuscope | LD_PRELOAD library that profiles GPU memory and kernels in unmodified programs by intercepting CUDA driver calls, at +1.6% overhead on a ResNet-18 training step. Allocator bytes and kernel launches per step match PyTorch exactly, with kernel time within 2.6% of torch.profiler. | C++, Python, LD_PRELOAD, CUDA, PyTorch | Code |
| Crime Scene Survey Robot | In a 2-person team, fine-tuned RT-DETR-L on a new 24,653-image, 8-class dataset (test mAP@0.5 0.79 on firearms and shell casings, 0.53 overall, at 43 FPS) and ran it on a real Unitree Go2 EDU quadruped, projecting LiDAR points into detections to place 3D evidence tags in a SLAM map. | PyTorch, RT-DETR, ROS 2, SLAM Toolbox, Nav2, LiDAR | Code |
| SyncForge | Real-time collaborative code editor. Yjs CRDTs converge concurrent edits without server-side conflict resolution; untrusted code in 6 languages runs in ephemeral Docker sandboxes with no capabilities or network access; WebSockets scale across Node.js instances with the Socket.io Redis adapter. | React, TypeScript, Node.js, Yjs, Socket.io, Redis, Docker | Demo · Code |
| QFlow | Event-driven trading strategy backtester with point-in-time data access (no look-ahead bias), slippage and commission models, Celery priority queues, and 19 risk metrics over 10 years of OHLCV data for 14 symbols. | React, FastAPI, TimescaleDB, Celery, Redis, Docker | Demo · Code |
| DriveLegal | Indian road traffic law assistant, built by a 6-person team where I was project lead and backend architect. The 2,400+ traffic laws and the fine calculator work offline in the PWA and Android app; the RAG legal assistant runs on the FastAPI backend and cites Motor Vehicles Act sections. | React, FastAPI, ChromaDB, Docker, Capacitor | Demo · Code |
| PhotoInsights | Serverless photo gallery on API Gateway and 6 Python Lambda functions, with direct-to-S3 uploads via pre-signed URLs, Amazon Rekognition tagging and OCR, and search by tag, text and date. | AWS Lambda, S3, DynamoDB, API Gateway, Cognito, CloudFront | Code |
- AWS Certified Solutions Architect – Associate (2026)
- Oracle Cloud Infrastructure Generative AI Professional (2026)
- IBM Data Science Professional Certificate (2026)
- Docker Foundations Professional (2025)

