KODAMA The Applied-AI Program
Kodama Mirai is an applied-AI company. Every product we build runs its AI on local model infrastructure — no third-party AI APIs. That is not a cost preference; it is the guarantee our regulated customers require: documents, questions, and answers never leave the environment the institution controls. It also means our research and evaluation work runs on hardware we own — which is why this page exists.
Four programs, one discipline
Production inference — the Workbench. The AI Compliance Workbench answers policy questions with validated citations from models running entirely on local infrastructure. Every answer records the model that produced it, frozen at the moment of generation. Evaluation sweeps against control-evidence corpora are how we keep that promise honest — and they are compute-hungry by design.
Agent evaluation — Kodama Forge. Our agent factory builds specialized IT-controls audit agents to written specifications, with a human review gate on every agent and a “specified, not shipped” discipline until each one ships. Before any agent reaches a customer it must survive evaluation harnesses that replay long-horizon work end to end — the same task, many runs, checked against the specification. Method commitments are published on the roadmap; the agents’ method is retrieval-first over your governed corpus.
Specialist-model research. Separate from the retrieval-first agents, we maintain a research track training specialist compliance models on a purpose-built corpus assembled from authoritative public sources. This is where fine-tuning hardware matters: the research goal is a model that speaks IT-general-controls natively, evaluated against the same evidence standard as everything else we ship.
Speech pipeline — Scam Tricker. Scam Tricker runs speech-to-text and text-to-speech locally, where call latency is the product: the pipeline must listen, decide, and speak inside a conversational pause. Latency work at that grade is measured on real hardware, not simulated.
Research infrastructure — current needs
| Program | Need | What it unlocks |
|---|---|---|
| AI Compliance Workbench | Dedicated inference/evaluation node: 2× RTX 6000 Ada-class GPUs (48 GB each), 256 GB RAM, 8 TB NVMe | Continuous evaluation runs against control-evidence corpora at production model sizes — the citation-validation guarantee, tested at full scale |
| Kodama Hub | Single-GPU workstation plus a staging server for ERM data pipelines | Anomaly-detection model research over risk-register and control data (research track; not yet a shipped feature), staged safely away from production |
| Kodama Forge — agent evaluation | 4-GPU training/serving node (H100-class, or 4× RTX 6000 Ada) | Multi-agent evaluation harnesses and long-horizon replay — enough parallel capacity to run agent cohorts against specifications before any agent ships |
| Specialist-model research | Fine-tuning node (DGX Spark-class; minimum one 48 GB+ GPU workstation), 128 GB RAM, 16 TB storage | Training and checkpointing specialist compliance models on the research corpus, on-premises end to end |
| Scam Tricker — speech pipeline | Workstation-class GPU test rig, plus telephony test numbers and devices | Local speech-to-text / text-to-speech latency testing at conversational grade, ahead of the real-call trial |
Each node serves a named program with a named outcome. Grant applications from Kodama Mirai reference this page; the specifications here and the specifications in any application are kept identical.