I build systems that compute on data they aren't allowed to see, and tooling that decides which vulnerabilities are actually worth fixing. Currently working on homomorphic-encryption inference with faculty at LNMIIT, and on LLM-assisted security tooling.
Omni Infoword Pvt. Ltd. · Cybersecurity Intern · May–Jul 2026 (~7 weeks) SOC alert triage across Splunk, Microsoft Sentinel, QRadar and Elastic — suspicious-login investigation, phishing analysis, EDR policy management. Second half on application security: OWASP Top 10 testing of pre-release software, including exposed staging endpoints and sensitive files reported to the dev team.
LNMIIT · Undergraduate Researcher · 2025–present Encrypted-inference and transciphering research with Prof. Debranjan Pal. Paper in preparation.
| Grand Finalist | PSB CyberShield 2026 — national hackathon, finale at IIT Hyderabad |
| HTB Academy | Skilled rank · working the Bug Bounty Hunter path |
| Certification | Google Cybersecurity Professional Certificate |
-
PPFDaaS — Fraud-detection inference over encrypted transactions. CKKS via Microsoft SEAL 4.1.2, C++17 gRPC server, Python client. The vendor never holds a secret key. Transciphering arm cuts client upload 249× by shipping a symmetric ciphertext evaluated homomorphically server-side. Digest-pinned containers, cosign signatures, SBOMs.
-
SetuGuard — Android malware triage joined to mule-account detection. Static APK analysis → RAG reporting → YARA generation, plus a gradient-boosted account scorer linked on certificate-hash indicators. Measured AUC 0.14 against a never-tuned-against corpus of real banking apps — the scorer ranks legitimate banks above malware, because both classes request the same permissions by construction. Documented rather than hidden.
-
reachability-triage — Ingests OSV/GHSA advisories and determines whether a CVE is genuinely reachable from a given codebase. Python, FastAPI, PostgreSQL, Docker, GitHub Actions. In development.
-
PwnBot — LLM pentest assistant over the Groq API. Structured tool-output parsing and a scope guard that refuses targets outside the declared engagement.
Every headline number in these repositories traces to a committed artifact file. Where a measurement turned out to be invalid — wrong parameter set, contaminated corpus, unreproducible host — the correction sits next to the original instead of replacing it. MEASUREMENT_PROVENANCE.md in PPFDaaS is the clearest example.
Working with: C++17, Python, Go, Docker, Microsoft SEAL, gRPC, FastAPI, PostgreSQL, GitHub Actions, Linux.