Blog

Running your own LLM inference is a function of five numbers most estimates miss: total parameters (not active), KV cache dtype, throughput bottleneck, replica count, and tensor-parallel topology. I built a calculator to get them right and benchmarked the results.

Most LLM tutorials show how to call a model. Almost none show how to make that call survive an audit. A seven-step compliance pipeline for MCP tool servers with RBAC, PII detection, policy enforcement, and structured audit logging.

Manual QA samples 2% of calls. Scrutiny scores 12 FDCPA/Reg F rules in 60 seconds with a dual-path evaluator. One LLM call for semantic rules, deterministic Python for metadata rules. Here is the architecture and where it breaks.

Phrase-matching covers 60-75% of call-center QA. The remaining 25-40% is where regulatory exposure lives. A well-designed LLM rubric grader closes the gap to 90%+. Here is the pattern for closing that gap with LLMs without hiding the limitations.

RegTriage is designed to train RL agents to audit contact center transcripts for compliance violations. Baselines across three open-source models show that Gemma 4 31B nearly solves the Hero Agent trap zero-shot. The RL training signal is still there, but it is weaker than we thought.