One engineer, three faculties.

It is not three tools. It is one virtual Kubernetes engineer, split across three ports only to keep each surface uncluttered. You converse with it; it acts on its own; and it gets measurably better at your cluster over time.

Fig 1 · Three ports, one engineer. Each faculty is independent but reads as one operator.

The differentiator is not the chat — plenty of tools chat about Kubernetes. It is that this engineer observes (9998), acts (9999), and retrains on its own incident history (9996), becoming a specialist in your cluster rather than a generic model. That closed loop is covered in The closed learning loop.

The console & sidebar.

The conversational front. You ask in plain English; every turn is grounded in a live cluster snapshot. The model answering is selectable — including the Factory's own finetunes.

§Chat — the cluster-aware agent

The Chat pane injects a live snapshot (nodes ready, pod phases, restart counts, pending pods, top-memory pods) ahead of your question, so answers reason from real signals. Type plain English, or click a slash-command card.

§The panes

§The model dropdown & per-model personas

The top-right dropdown is provider-aware: it probes Ollama, LM Studio, Claude Max, LiteLLM — and the Model Factory (9996/v1/models). Every finetune the Factory serves shows up as a distinct entry under the <domain>-sre naming convention.

Selecting a Factory model swaps both the system prompt and the chat cards:

Tabs & the stable endpoint.

The model lifecycle, decoupled from the SRE: build a dataset, finetune it on burst GPU, serve the result on a stable endpoint the SRE consumes. Re-finetuning hot-swaps the weights — the URL and the dropdown entry never change.

§The stable, multi-alias endpoint

# one URL, many models, never re-URLs on retrain
GET http://hub:9996/v1/models # → k8s-sre (default), insurance-sre
POST http://hub:9996/v1/chat/completions {"model":"k8s-sre", "messages":[...]}

§Datasets

Three sources feed one registry:

a.Synthetic

An LLM (OpenRouter, Ollama fallback) turns a plain-English description into a structured CSV. Controlled, repeatable — ideal for demos.

b.BigSet (real web data)

Self-hosted TinyFish BigSet (ports 3500/3501) researches the live web and returns a verified, source-attributed table. Used to build the us-health-insurers set (15 real insurers).

c.SRE-history

The cluster's own record — /api/history events plus daily audit summaries and published reports — converted to training rows. This is what makes the loop close.

§Finetune — burst, persist, serve

Fig 2 · The finetune pipeline. Engine: transformers + peft (Unsloth deferred — pip overlay broke torchvision on the generic CUDA image).

§Scheduler — keep the model fresh

The Scheduler owns the retrain knobs: alias, dataset (dropdown from the registry), base model (swap to Qwen/Qwen2.5-1.5B-Instruct when the corpus is rich), and cadence (manual / 6h / 12h / daily / weekly) with an Enabled toggle. Each fire rebuilds+enriches the corpus, finetunes, and hot-swaps behind the alias.

Autonomous SRE & Telegram.

The faculty that acts on its own and reaches your phone: scheduled audits, report generation, and a two-way chat surface.

§Autonomous SRE

Hermes drives scheduled work via cron writers (daily-audit.sh, daily-report.sh): a deterministic cluster snapshot, summarized by the Console's model, written to the audit log + a saved HTML report, and pinged to Telegram.

§Two bots, two directions

§Commands

Plain English routes to the cluster-aware agent. Slash commands map to the active model's cards. And:

/retrain # retrain the default alias
/retrain k8s-sre # retrain a specific model — hot-swaps behind its alias when done
/assess_health /triage_restarts /disk_pressure /daily_audit

The closed learning loop.

Most assistants are read-only or generic. This one feeds its own operational record back into training — so it specializes in your cluster, and improves every retrain.

Fig 3 · The loop. New incidents accumulate → the Scheduler retrains → the model gets sharper, behind the same alias.

Procedures — how do I…

Ports, deploys, env & API.

§Port map

§Deployment matrix

§Env vars

§API index (Model Factory)

Troubleshooting.