Ram Charan Polisetti

DBA candidate in Applied Artificial Intelligence, Belhaven University. I build open-source tooling for AI governance — bias audits, red-teaming, evidence bundles, approval registries. Indianapolis, IN.
ramcharansatyasaitej@students.belhaven.edu · GitHub

Projects

Open-source AI governance tooling. Everything is public on GitHub, tested, Apache-2.0.

vendor-diligence-kit (2026)
Standardized diligence for third-party AI systems: vendor intake with hashed claims, a test battery (red-team + fairness + lineage) that delegates to pinned sibling tools rather than reimplementing them, a weighted risk score, and a go/no-go written into the governance registry as the system's first approval record. In the demo, a deliberately weak vendor scores 47.1 (rejected) and the remediated version scores 95.0 (approved). 19/19 tests.

governance-evidence-vault (2026)
Seals everything behind a deployed system — conformance packs, registry records, incident reports, evals — into one signed, hash-chained bundle an external auditor can verify offline, with no tooling beyond the standard library. A tampered copy is caught and reported explicitly. 31/31 tests.

opsaudit (2025–2026)
Bias and disparity audit toolkit: the core audit engine (64/64 tests), a deployment gate that blocks biased models before they ship (86/86 tests), and a synthetic-data auditor (119/119 tests). The repo's own disparity gate runs on its PRs.

garak (2026)
Fork of NVIDIA's LLM vulnerability scanner with a real bug fix: Buff._derive_new_attempt passed notes and detector_results by reference, so sibling attempts shared mutable score state. Fixed with shallow copies. 5 new tests, 42 regression tests pass.

ai-incident-runbook (2026)
Runbook and lifecycle tracking for AI incidents in production. 40/40 tests, with a verified five-event incident lifecycle.

More governance tooling:

eval-judge-calibration — how well LLM judges agree with human labels before trusting them in evals (46/46 tests; Cohen's kappa 0.769, single-annotator labels documented as a limitation).
rag-eval-drift — detects RAG answer drift as documents change; removing two policy docs dropped groundedness 100% → 84.2% (16/16 tests).
conformance-pack — structured conformity evidence for high-risk AI use cases (54/54 tests; clean hiring model → conditional, biased model → no-go).
rag-redteam — red-teaming harness for RAG pipelines; 4 of 6 poisoned-chunk attacks succeeded in the live demo (45/45 tests).
regulatory-change-watcher — watches AI regulation changes and flags what affects registered systems (21/21 tests).
ai-act-checker — maps a system to EU AI Act risk tiers and deployer duties (48/48 tests).
disparity-monitor — disparity and bias metrics for model outputs (55/55 tests).
model-governance-registry — model records, approvals, and a stale-approval sweeper (54/54 tests after integration).
rag-governance-demo — the first build: RAG demo with governance wired in (41/41 tests).
rai-monitor — responsible-AI signal monitoring on live systems (104/104 tests).

Earlier / experimental:

Q3_Hackathon — quantum hackathon work; two files licensed to Yonsei's HAIR Lab for their Parallel Quantum Instruction Dataset.
QuantumAlgorithms, oilspill, Agent-Zero, BendTheCurve, Expressive-Resume — earlier experiments.

About

I work on AI governance from the operator's side. Before the doctorate I spent years in supply chain analytics and applied AI — transportation network optimization at Amazon, then AI solutions and data analytics consulting at Cortracker360 (2025–2026): RAG/LLM platforms, BI and data engineering at scale. Now I'm building the unglamorous parts of governing AI systems: bias audits that run as deployment gates, red-teaming harnesses, evidence bundles an auditor can verify offline, and registries that record go/no-go decisions.

Interests: responsible AI systems and pre-deployment governance; bias and disparity auditing in operational decision systems; AI policy, risk tiers, and deployer duties (EU AI Act, NIST AI RMF); red-teaming and evaluation of RAG/LLM systems; evidence, audit trails, and incident response for deployed models.

Writing

The Gate That Says No (draft, in review) — essay arguing bias audits arrive after the damage, so the check should sit in the deployment pipeline as a gate that can say no.
DBA research (in progress) — governance of AI in operational decision systems.

Contact

Email: ramcharansatyasaitej@students.belhaven.edu
GitHub: ram-polisetti