CSE (Cyber Security) @ VIT Vellore · Class of 2028
I like finding out how systems break, and then building the thing that notices. Most of what I build sits somewhere between security, machine learning, and hardware that isn't fast enough to run the model you wanted to run.
Open to Summer 2027 internships — software engineering, security, or applied ML. Based in Vellore / Bangalore, India. Happy to talk remote.
🔍 Rubble Runners — Smart India Hackathon · AI/ML Engineer An edge node for search-and-rescue: an ESP-CAM goes into collapsed rubble, a Snapdragon phone stays outside as a base station and runs person detection on the incoming stream, then saves positive detections with coordinates. I own the ML end — dataset prep and annotation strategy, training a lightweight YOLO variant on CrowdHuman, evaluation, and the TFLite / Qualcomm SNPE conversion needed to hold ~5–10 FPS without cooking the phone.
The interesting constraint isn't accuracy, it's thermal budget. You can't just run a bigger model. Also worth saying plainly: CrowdHuman mAP is not search-and-rescue performance, and I'm treating evaluation readiness — not dataset size — as the real bottleneck.
🛰️ Autonomous drone mapping — Team Ardra, Software Domain
Real-time aerial mapping on a Jetson Orin Nano with TensorRT. Three pieces:
MapNode (SuperPoint + LightGlue feature matching), MLNode (GPU preprocessing,
async writer, three-thread pipeline over a spatial GPS grid), and MapStitcher
(dynamic canvas — most of my time there went into a boundary-mask fix for a
speckle/banding artifact at tile seams).
🏃 Borderline (WIP) — Strava, but competitive. Turning GPS traces into contested territory instead of a leaderboard you scroll past once. Currently at MVP; the hard parts are trace processing and making cheating unrewarding rather than impossible.
| Project | What it is | Stack |
|---|---|---|
| SIEM-Lite | A miniature SIEM: log ingestion, parsing, correlation rules, and alerting. Built to actually understand what commercial SIEMs are doing rather than to click through one. | Python, Linux |
| VTOP GPA Calculator | GPA/CGPA dashboard for VIT's student portal. Built it because the official one doesn't let you plan — now other students use it. | JavaScript |
| Cosmopedia | Solar-system encyclopedia with Firebase auth, Firestore favourites, canvas animation and NASA APOD. Started vanilla JS, migrated to React + Vite deliberately, one layer at a time. | React, Vite, Firebase |
| Rooftop detection from orthophotos [TODO-LINK] | YOLOv8 on tiled aerial GeoTIFFs — rasterio tiling, shapefile→YOLO label conversion, four roof classes with severe class imbalance. Basis for a co-authored research paper. | YOLOv8, rasterio, PyTorch |
| Hybrid file encryption [TODO-LINK] | AES for the payload, RSA for the key. Small tool, but it's where I learned that "it works on my machine" and "the keypair regenerates every run" can be true simultaneously. | Python, cryptography |
| Keylogger (educational) | Input-capture research tool built during my Pinnacle Labs internship, for understanding endpoint telemetry and detection. Scope and limitations documented in the repo. | Python |
| Honeypot on Raspberry Pi [TODO-LINK] | Low-interaction honeypot on a Pi to see what actually knocks on a residential IP. Spoiler: constantly, and mostly for root. |
Linux, Python, RPi |
| Portfolio | Personal site — projects, writing, and the long version of all of this. Source | HTML/CSS/JS, Vercel |
Probabilistic rooftop detection from aerial imagery — co-authored with Kriday Narula, faculty-supervised. IEEE format. Grew out of the GeoTIFF/YOLOv8 pipeline above.
Nuclear thermal rocket design — co-authored, faculty-supervised. Not my domain by training, which is exactly why it was worth doing.
Team Ardra — Software Domain · Sept 2025 – present Autonomous drone systems R&D. Real-time mapping and on-board inference (see above).
Pinnacle Labs — Cybersecurity Intern Endpoint security tooling and input-capture research.
I've tried to be honest about the tiers here, because a stack list nobody can defend in an interview is worse than a short one.
Reach for daily — Python JavaScript C Linux Git Bash
Comfortable — React FastAPI Flask Node.js Docker PyTorch OpenCV NumPy Pandas Firebase
Actively working in, still learning — TensorRT TFLite Qualcomm SNPE Jetson Qdrant TypeScript C++
Security — SIEM design & log analysis, honeypots, endpoint telemetry, applied cryptography (AES/RSA), Linux hardening
Poetry, chess, the gym, and a 19th-century novel I keep not finishing. Currently planning a Hampta Pass trek, which is the only pipeline I have with no error handling.
Genuinely open to conversations — internships, project collaborations, or someone telling me my approach to something is wrong.
📧 raj.tibarewala@gmail.com · 💼 LinkedIn · 🌐 raj-tib.vercel.app