Autotune proposes the next run's hyperparameters & mode (NN warm-start → RK4 sweep), routes queued content into batches, and ingests it. Heavy neural retraining runs off-box on the GPU rig using this config.
Recent runs
when
strategy/cfg
added
chunks
📥 Content queue
Upload/queue material for the model to learn. It's held until a training run consumes it.
Pending
subject
name
pri
chars
by
🔑 API keys
Generate tiered keys to share access. low=query model/KB · medium=+queue content · admin=+train/manage. Keys are shown once and stored hashed.
label
tier
bits
id
status
🗝️ API key vault
Store third-party API keys (other sites) encrypted at rest. Reveal only when needed.
name
stored by
👥 Users & access
Grant portal access. You are the permanent sudo super-user.
user
role
📚 Dataset importers
Generate training items straight into the queue. The third is a modular sandbox — paste anything and choose how it's parsed, to test new sources.
🛡️ Guardrails
Add, modify and monitor the safety filters. Toggle categories, tune severity, add regex patterns, and test live. Changes apply immediately and are logged.
category
severity
on
patterns
Recent guardrail hits
when
kind
detail
sev
🎛️ Parameters
Inspect and override the NN+RK4 search bounds, and keep your own typed knobs to test new things. The autotuner respects these immediately.
Autotune search space
param
lo
hi
scale
name
value
type
note
🧠 Training & GPU
Keep the model busy. The idle collector mines fresh skill data into the queue so it's never sitting idle; skill routines target specific capabilities; heavy retraining is dispatched to your home GPU box, which pulls jobs, trains, and reports back.
🧠 TriAi Mind — personality & cognition
Tune how TriAi thinks and talks. Traits shape its tone and manner; cognition shapes how deeply it comprehends, how much it remembers across sessions, and how far it reaches for out-of-the-box ideas. Saved live to the running model.