Cybersecurity undergrad building systems close to the metal — encrypted storage, boot integrity, and AI-driven security tooling.

I build security systems (filesystems, boot chains, crypto) and AI/RAG tooling for security operations, mostly in C/C++ and Python.

• Qwen-ATLAS — RAG-based threat intelligence agent (team project, Model Architect role)

C OpenSSL libfuse3 libargon2

A block-level encrypted virtual filesystem built on FUSE, verified against 7 distinct attack classes (ciphertext tampering, cross-container block transplant, block reordering).

• Per-block AES-256-GCM encryption with CSPRNG nonces — concurrent pread/pwrite access without full-file lock contention
• Two-key derivation chain (Argon2id → HKDF-Expand, domain-separated labels) enforcing cryptographic isolation between containers
• Validated under simultaneous read/write access using GDB and Valgrind (zero leaks across ~1,000 allocations)

Known limits: no replay/rollback protection, no crash consistency (documented, not hidden — Phase 2 addresses this)

🔗 Repo · Tagged v0.1.0, MIT licensed

Python ChromaDB HuggingFace LoRA

Hybrid RAG pipeline over 915 MITRE ATT&CK STIX objects for cyber threat intelligence retrieval and attribution.

• Raised threat attribution accuracy from 43.75% → 83.75% on a 40-query frozen benchmark
• Deterministic benchmark across 8 threat intel categories with a reproducible scoring rubric
• Adversarial evaluation: measured nation-state misattribution rate under poisoned inputs

🔗 Repo

Python FastAPI React NetworkX Llama-3

AI-driven DevSecOps platform that parses Terraform into DAGs and simulates cascading "blast radius" failures.

• BFS-based blast-radius simulation over infrastructure dependency graphs
• Chaos simulations offloaded to a self-hosted Llama-3-70B model on an AMD MI300X GPU (via AMD hackathon access), with Fireworks API fallback for logic synthesis
• Custom GitHub Action posting AI-synthesized SOC2 review comments and HCL patches directly on PRs

🔗 Repo

📫 Reach me at karthiksethuraman6@gmail.com or on LinkedIn