kbird · Kristopher RichardsI find how your AI fails
before your users do.
kbird is the independent practice of Kristopher Richards: LLM red teaming and evaluation for teams shipping models and agents. Remote, US-based, contract or project work.
Services
Adversarial testing that finds real failure modes, with evidence your engineers can reproduce.
LLM & agent red teaming
Jailbreaks, direct and indirect prompt injection, tool misuse, data exfiltration, and multi-turn drift.
Evaluation harnesses
Versioned adversarial suites, deterministic scoring, tamper-evident transcripts, and reports your team can rerun.
Bias & reasoning audits
Graph-based detection of bias, sycophancy, and unverified claims in agent memory and model outputs.
Private model testing
Open-weight models tested on local hardware, so your prompts and data never leave the box.
Selected work
Built and run end to end. The harness and Debias are open source.
LLM Red-Team Harness
69 versioned attack cases, deterministic scoring, and hash-chained transcripts. Indirect prompt-injection runs caught canary-token leaks in gemma3:4b and Qwen3-8B.
Read more →Agent memory · Graph Hacks 2026Debias
Memory you can audit: a bias-aware memory graph for AI agents on FalkorDB, with five pure-Cypher detectors and no LLM judge.
Read more →Applied AI · SMB operationsMarket Leaks
A profit-leak finder that shows which small-business orders lost money the moment they were accepted, reconciled to the cent.
Read more →Why work with me
Years in safety-critical rail operations, where a missed signal is not a bug report. I bring the same habits to AI: written procedure, full logging, and assuming failure until it's proven otherwise.
Let's break something on purpose
Shipping a model, an agent, or a RAG system? Tell me what it can touch and I'll tell you how it fails.