Tinman Lab

Autonomous Machines. Second-Order Systems.

AGENT MEMORY · ADVERSARIAL SAFETY · AGENTIC ECONOMY · PERCEPTION SYSTEMS · APPLIED RESEARCH

We build on-device AI systems that reason, remember, and self-correct — small models designed to run autonomously at the edge with calibrated uncertainty and adversarial robustness.

Research Areas

Agent Memory — Encrypted semantic memory infrastructure for persistent agent context
Adversarial Safety — Multi-agent stress-testing and trust verification for autonomous systems
Perception Systems — On-device vision, voice, and multimodal understanding
Disposition Distillation — A three-arc study finding that imitation, attention-head tempering, and frozen-base sidecars all fail to move judge-measured disposition without damaging content quality at sub-billion scale (arXiv:2604.11867).

Open-Source Releases

Tinman SmolOmni (MLA) — small omnimodal models with multi-head latent attention.

Model	Description
Tinman-SmolOmni-MLA-256M	256M parameter omnimodal
Tinman-SmolOmni-MLA-500M	500M parameter omnimodal
Tinman-SmolOmni-MLA-Toolkit	Training and inference toolkit

Tinman Companion — Gemma 4 fine-tunes for on-device companion use cases.

Model	Description
Tinman-gemma4-companion-merged	Full-precision merged model
Tinman-gemma4-companion-gguf	GGUF quantized for llama.cpp
Tinman-gemma4-companion-litert-lm	LiteRT-LM for on-device deployment
Tinman-gemma4-companion-sft	SFT checkpoint
Tinman-gemma4-companion-dpo	DPO checkpoint

Links

Website GitHub Paper (arXiv) DD Artifacts