Chandhan Saai Katuri

Designed and shipped a vision-LLM audit pipeline that reads contractor screen recordings against per-project task policies and seals priced, recomputable audit ledgers.

Holds up 80% precision and recall on the golden evaluation set; 264 sessions sealed in production. Read models benchmarked across providers before one was chosen.

Architected the core of an LLM data-annotation platform: five-role access control, an audited workflow state machine, an append-only audit log.

Holds up Audit coverage proven across all 54 endpoints; test suite grew 210 → 1,239. Django REST Framework and React.

Delivered a finance-interview rubric verifier in six days: LLM-judge claim validation, shadow-verifier gating, hard execution limits.

Holds up Set the team standard that a test only counts once a deliberate source defect proves it can fail, and the defect itself is verified as landed.

Led SOC 2 Type II readiness during an active audit window, including a pgAudit and storage-encryption migration across 11 production Postgres databases.

Holds up Control coverage moved 154 → 165 of 173 (95%), with 40+ evidence artifacts and penetration-test findings remediated.

Built a multi-scanner security pipeline with LLM-assisted triage across five static-analysis engines.

Holds up 345 raw findings reduced to 33 verified, nine of them critical: remote code execution, SSRF, IDOR, exposed credentials. Pipeline is public on GitHub.

Evaluated frontier image-generation models against calibrated quality rubrics to produce preference data for RLHF pipelines.

Holds up Rubrics covered rendering artifacts and anatomical consistency; also reviewed ML codebases for algorithm selection and implementation quality.

Selected as a Star Fellow on a program probing model reasoning boundaries.

Holds up The work ran against frontier research publications, testing where model reasoning breaks down.