9 projects
outside-in-alignment
Operator-side discipline layer for LLMs. Compensates for calibration, memory, sycophancy, and action-feedback gaps.
adaptmem
One-shot domain adaptation for retrieval — beat your baseline with 200 lines of hard-negative mining.
mnemonics
Local-first AI memory — semantic retrieval over your own SQLite + HNSW index.
nakata-agentguard
Lightweight agentic loop detector and safety monitor. No LLM required.
nakata-promptguard
Prompt injection / jailbreak detection — the input-side gate of the no-LLM-judge stack.
nakata-truthcheck
Open-world fact verification for AI claims (sibling to halluguard, which handles closed-world).
freshmint
Pythonic C2PA, sign every AI-generated image with verifiable provenance.
halluguard
Reverse-RAG hallucination detector. No LLM judge. Vendor-neutral.
claimcheck
Domain-tuned retrieval + zero-LLM claim verification in one pipeline.
