Security engineering & data analytics. I build things I can explain in depth, then measure them until they tell me something I didn't expect.
🌐 Portfolio: nishankswamy.github.io · finishing an MSc in Cyber Security in Berlin.
I recently ran a 30-day build challenge:
ten projects across security and data, each shipped with a benchmark, tests, CI,
and a written finding that contradicts the obvious assumption. 831 tests · CI on every repo · 2 capstone platforms.
The rule for every project: build the core, do the hard half, then break it and write down the result. A few that came out of that:
- Forged a MAC without the key.
SHA256(secret‖message)is length-extendable — a digest is the hash's internal state. Appendedrole=adminto a signed message in milliseconds. → applied-cryptography - A benchmark that lied by 600x. A "678x cache speedup" was timing the error-fallback path; the real number was 1.1x. Now I distrust results that flatter the design. → url-shortener
- ML lost to five lines of statistics. Isolation Forest scored worse than an STL+MAD baseline on security telemetry at a realistic base rate. "Use a model" is not a strategy. → anomaly-detection
| What it is | The finding | |
|---|---|---|
| 🛡️ SIEM platform | Multi-source ingest → inline detection + behavioural baselines → entity-first investigation | Detects a full attack campaign in 13ms; alert → evidence is one indexed lookup, 26x faster than a scan |
| 📊 Analytics platform | Quality-gated ingest → star schema with cubes → self-serve queries with a cost display | What breaks at scale is memory, not compute — 530 B/row as dicts vs 50 on disk |
Cryptography · detection engineering · a columnar query engine · anomaly detection · network traffic analysis · an exactly-once streaming pipeline · differential privacy — plus the two capstones. Everything is defensive; the only things I attack, I attack in my own code.
→ github.com/nishankswamy/30-days — the full index, findings, and case studies.
📧 Nishankde@gmail.com · 🌐 Portfolio · 💼 LinkedIn