Natural language, with explicit cost gates and a sealed transcript.
The assistant drives every product on the platform — datasets, synthesis, SCADA, ICS, agent runs — through plain English. It plans first, asks for approval, streams the execution, and seals the whole conversation into the same evidence chain as the rest of the stack.
- Tools wired
- 43
- Tool families
- 5
- Cost gate
- per step
- Transcript
- sealed
Six rules the assistant always follows.
The same rules whether you're asking it to clean a CSV or replay an ICS attack scenario. No hidden tool calls, no silent retries, no surprise spend.
Every one of the six rules below exists because of a specific failure mode large-language-model agents have in production when they are wired to spend money or touch production data. Rule 01 (plan-first, execute-second) is there to close the "agent hallucinates a tool call and the tool succeeds anyway" gap. Rule 02 (cost-gated by default) is there to close the "agent fans out 500 calls before the operator notices" gap. Rule 03 (job envelope) is there because multi-step runs without an envelope become long-lived zombie jobs that nobody remembers started.
Rules 04 through 06 close the trust gap between the team running the assistant and the team auditing it: the tool surface is typed so there is no ambiguity in what an action did, the evidence chain is sealed so there is no ambiguity in what an action produced, and every transcript is replayable so an auditor can reconstruct a decision without calling the analyst who made it.
The commercial cut: an assistant that a procurement lead can approve for credit-card-grade spend without needing an engineer to custody it. Every rule is a control that already exists; every rule is a reason the customer does not need to bolt their own control on top.
Plans first, executes second
Every multi-step task starts with an explicit plan card. No blind tool calls — the operator approves before any credit is spent.
Cost-gated by default
Each step carries a credit estimate; the assistant pauses when a single step exceeds your soft cap.
Live exec stream
Tool runs stream in real time — start, progress %, output preview, and per-step duration.
Cancel and refine, mid-run
Stop a plan part-way, refine the prompt, replay from the last checkpoint. No work re-done unnecessarily.
Same evidence chain
Every artefact joins the same cryptographic hash-chained ledger as Mock, Synthesize, SCADA, and ICS.
Replayable transcripts
Every conversation seals into a transcript artefact: prompt, plan, approvals, tool I/O, signed bundle.
Forty-three tools the assistant can wield.
Five families, all typed, all callable from natural language and from the SDK. The assistant chooses; you can override.
Data prep
Machine learning
Explain & report
Synthetic data
Industrial / ICS
Built to survive procurement & security review.
Tenant-isolated
Conversation, prompt, dataset preview — never leave your VPC.
Cost-bounded
Soft-cap per turn, hard-cap per project, both visible to operator.
Pausable
Stop mid-plan, resume from last sealed checkpoint. No re-run.
Auditable
Every prompt, plan, approval, tool I/O joins the cryptographic hash chain.
Honest answers to the four questions everyone asks.
If yours isn’t here, the working session is the fastest way to a real answer.
Which model drives it?+
Auto-routed: a small open model proposes the plan, a larger model only fires when reasoning depth justifies the cost. You never pick the model directly.
Where is my data sent?+
To your tenant's compute pool only. Prompts, dataset previews, and tool I/O never leave the RadMah AI VPC. Air-gapped deployments are available on Enterprise.
Can I disable it per role?+
Yes — RBAC scopes (read / generate / admin) gate which tool families a role can invoke. Audit log records every call.
What if a step fails?+
The assistant retries deterministically inside its budget, then surfaces the exact failure to you with a recommended fix. No silent fallback.
Bring a real prompt. Watch the plan card.
A thirty-minute working session. You bring a representative dataset (or we mock one to match your schema) and a single question your team actually needs answered — fraud-score on a held-out tranche, forecast the next quarter of transactions, profile the drift between Q1 and Q2. We drive the assistant end-to-end while you watch the plan card assemble itself in real time, approve the spend, and walk away with the sealed transcript bundle the auditor can replay.
For the data-team lead: this is how you demonstrate to your VP that AI orchestration is ready for your stack without committing to a six-month implementation. For the procurement lead: cost-gated by default, no runaway spend possible, every credit line-itemised in the transcript. For the compliance officer: every tool call versioned, every approval logged, every output signed into an evidence bundle a regulator can open offline.
You leave with four concrete artefacts: (1) the sealed plan card the assistant generated, (2) the sealed transcript of every tool call, (3) the generated artefact itself (dataset, report, forecast) with its quality report, and (4) a one-page summary your exec sponsor can take to the steering committee without needing to learn the SDK.