Helm. The two-sided LLM. The human still has the wheel.
An LLM on its own is an opinion factory. Helm is the layer that lets a coach, a reviewer, a clinician, an admissions officer steer the model in real time — nudge it before it commits, rewrite a sentence before it lands, hold it to a posture you can defend out loud. This is our enterprise-distinctive surface, and the work the rest of the platform is built around.
Two seats at the same conversation.
Most chat systems are one-sided: the model speaks, the user reads. Helm gives the conversation a second seat — a quiet operator window where a human can watch the response form, intervene mid-stream, and shape the next turn before it ships. The user sees a single, coherent reply. Behind it, a coach has been at the wheel.
Four primitives
- Nudge. Send a hint as the model is composing. “Pull back on the certainty,” “name the trade-off,” “ask a question instead.” The model adjusts before its next token.
- Guardrail. Hard rails attached to the conversation: never recommend X, always cite Y, refuse Z. Violations are caught before the user sees them, with the rule logged.
- Rewrite. Step into a sentence mid-stream. Edit it, replace it, or rephrase it. The user sees the corrected version; Observatory keeps the diff.
- Handover. Promote a turn from the model to the human, or back. Used when the question deserves a person, or when a specialist needs to take over a thread.
Three shapes, one engagement model.
Helm is offered as enterprise engagements only. Pricing is part of the contract; volume curves and committed-use discounts are posted on request.
- Sized for small operator teams
- Nudge, guardrail, rewrite
- Standard latency budget
- Observatory traces included
- Concurrent operators at scale
- Handover & supervisor escalation
- Custom rubrics & refusal posture
- Enterprise governance options on request
- Private deployment, your region
- Retention shaped to your policy
- Custom integrations, dedicated support
- On-call review of escalation protocols
One websocket per conversation. Two surfaces.
# Open a Helm conversation
POST /v1/helm/conversations
{
"user_id": "ash-14",
"harness": "default",
"guardrails": ["no-medical-advice", "cite-on-claims"]
}
# Operator stream (the second seat)
WS /v1/helm/conversations/:id/operate
> { "type": "nudge", "text": "soften the certainty" }
> { "type": "guardrail", "rule": "no-recommendation" }
> { "type": "rewrite", "patch": "...replacement..." }
> { "type": "handover", "to": "operator" }
# User stream (what the user sees)
WS /v1/helm/conversations/:id/talkThe difference between an LLM and a colleague.
An LLM in a vacuum is something between a vending machine and a courtroom witness — confident, fluent, and answerable to no one. The work that makes the model useful inside a real organisation is the work of someone who has seen this kind of thing before, watching the response form, and adjusting. Helm makes that work native, observable, and shareable.
We built it because we needed it. We use Helm to coach the model through admissions conversations with teenagers, to keep replies grounded when parents write in worried, and to hold the line on compliance, brand, and the way we want to sound to a parent — through conversations we cannot script in advance. None of that is possible with a one-sided API.
Put a human back in the loop without slowing the loop down.
Helm is offered as enterprise engagements only. Tell us about the workload — the people, the posture, the stakes — and we will tell you whether Helm is the right shape.