Raghav Pathak

• Ran an OWASP Top 10 test pass over a pre-release internal application, uncovering publicly reachable staging pages and exposed sensitive files before production release.
• Traced every finding to the developers who wrote the code, walked them through the exploit mechanism until each issue closed, and documented the 7-week engagement in the formal internship report.
• Triaged suspicious-login, malware and phishing alerts across 4 SIEM platforms (Splunk, Microsoft Sentinel, QRadar, Elastic) and ran Nessus and OpenVAS vulnerability scans.

Solo Submission Razorpay AI Buildathon 2026, Track 02 (AI Risk Manager)

• Designed an LLM responder that ingests payment chargebacks, assembles supporting evidence from transaction context, and drafts the dispute response an operations team writes by hand.
• Delivered it solo in 5 days with the model inside the decision path rather than generating text at the end of a pipeline; built a two-stage grounding gate ensuring zero fabricated reason codes or citations.
• Shipped as a public reproducible repository with strict evaluation harnesses and a 5-minute recorded demonstration.

Grand Finalist PSB CyberShield 2026, National Finale at IIT Hyderabad (Team of 4)

• Engineered a retrieval-augmented pipeline over a 95-package Android corpus: static features drive FAISS retrieval, then a local Mistral 7B emits a schema-constrained JSON report with MITRE technique mappings and generated YARA rules.
• Eliminated fabricated citations by restricting the schema to a fixed set of valid chunk IDs, plus a grounding gate that caught real hallucinations in live output, holding zero failures across 7 hostile-input cases.
• Pinned generation to temperature 0 and a fixed seed, making every report reproducible run-to-run.
• Accelerated the dataset-scoring endpoint 11× (7.6 s to 0.66 s) and cut peak memory 2.9× (1.9 GB to 0.66 GB) by lifting 5-fold cross-validation into a precomputed artifact.
• Audited the classifier across 9,082 accounts, isolated label leakage in graph features, and replaced a single-split score with a 20-seed repeated holdout (median AUCPR 0.271, AUROC 0.872).

• Implemented an agent that converts natural-language goals into sequences of tool calls, parsing raw command-line output into typed objects before the model reasons over it.
• Enforced a scope guard blocking any action outside an explicitly declared scope, and refactored into a multi-module package across 5 refactor phases with externalized configurations.

• Developed an inference service scoring fraud risk on bank data the server never decrypts, using CKKS homomorphic encryption; the model reaches 0.979 AUC on the ULB credit-card dataset.
• Compressed per-transaction upload 249× (262 KB to 1 KB) with a Go stream-cipher layer and benchmarked a 4.98× mean speedup over Lattigo at matched ring parameters (N = 8192).
• Released it as an 8-stage container build with digest-pinned images, cosign signatures and a per-release SBOM.