Verifiable output. Deterministic grading. Traceable failure modes.
Currently shipping: 30-day sprint. Four artifacts across the verifiability stack. New repo and post weekly.
- audited-tool-mcp — A compliance-aware MCP server. Seven-step pipeline: RBAC, PII detection, policy enforcement, audit logging. 15-case eval, 100% pass.
- RegTriage-OpenEnv — An OpenEnv RL environment where the reward signal is auditor approval. 12 tasks, severity-weighted F1, auto-fail caps.
- rubric-grader-eval — A reference pattern for compiling unstructured rubrics into evaluable schemas. Handles clean CSV, boolean composites, document masquerades.
Previously at Yactraq Online Inc.: shipped three LLM features to enterprise production — automated quality scoring, real-time compliance assistant, conversational analytics engine. Self-hosted on customer infrastructure. Regulated-industry deployment.
rituraj.info — notes on production ML, compliance systems, and agent architectures.
Open to short engagements: verifiability audits, eval-harness builds, fractional engineering for agentic systems in regulated domains.