AI systems · automation · engineering

I build AI systems that run, and know when to stop.

Agents, automations and software for research, monitoring and operations, with evidence attached and a human approving the steps that matter.

Decision rail

one proposal · one gate
Approved

Reviewable file

The proposal crosses the human gate.
Stopped

Output: —

Reason recorded: evidence is incomplete.
A proposal reaches a human approval gate. If approved, it becomes an inspectable output. If stopped, the output stays empty and the reason is recorded.

Work · selected systems

Claims on the left.
Proof on the right.

Each system is shown by the decisions it encodes and the evidence it leaves behind.

01 · Lead case

One working memory, many agents

ProductionInternal system

Claude, Codex and ChatGPT read the same work history instead of starting from zero.

Read the case
Working memory · authority and boundariessanitized policy diagram
  1. identity and permanent boundaries
  2. confirmed profile and preferences
  3. current confirmed decisions and open work
  4. original project documents
  5. the working-memory graph · history, not orders
Working memory

History and context, never orders.

Agents may query it; they may not refresh it or use it to skip approval.

Read-only for agents
02 · Gate case

AI Control Tower: delegation behind a gate

In developmentMCP server

Premium agents keep the reasoning. Bulk work goes to cheaper models, only after an exact approval.

Read the case

Method · seven rules

A short index of
non-negotiables.

  1. 01

    Evidence first.

    If it isn’t in the source, it doesn’t get written. Every claim, number and recommendation points back to a source. Gaps are labelled as gaps, never filled with a plausible guess.

  2. 02

    A human decides.

    Agents detect and propose. They don’t publish. Publishing, changing a live site, spending money or sending a message is a human act, every single time.

  3. 03

    Fail closed.

    When in doubt, the system stops instead of guessing. A missing file, an unknown state or an unverified claim is a stop with a reason, not a warning that scrolls by.

  4. 04

    Deterministic where possible, LLM where needed.

    A model never does what a script does better. Parsing, counting, comparing and formatting are code. The model gets the steps that genuinely need judgment, and nothing else.

  5. 05

    Privacy by design.

    Collect less. Protect what remains. Keys stay in the keychain, memories stay on disk, and personal data is redacted before anything is indexed.

  6. 06

    Write down the traps.

    Every mistake paid for becomes a rule for the next agent. The same trap is never paid for twice: it goes into the instructions the next agent reads first.

  7. 07

    Every output is inspectable.

    Versioned, reviewable, verifiable. Reports, plans and digests are files you can diff and review. If you can’t inspect it, you can’t approve it.

Read the full method

Proof · authentic artefact

The interface is a contract.

This excerpt comes from the actual AI Control Tower interface. It names the tools, the privacy classes and the stop conditions, without exposing a real plan identifier.

See the system around it
AI Control Tower · the six toolsinterface · authentic, sanitized
  1. ai_provider_statuschecks locally that credentials exist; never shows keys
  2. ai_plan_jobnon-executing plan: provider and model pinned, finite limits,
  3. and an exact approval phrase
  4. ai_submit_jobruns only if the phrase matches exactly
  5. ai_job_statusreads local state
  6. ai_get_resultnormalised results with provenance; partial is not complete
  7. ai_cancel_jobstops future work, keeps what was already created
  8. privacy classes
  9. publicexplicit or automatic routing
  10. privateonly after the provider data policy is assessed for that job
  11. sensitiveprovider must be named; stop if retention or region is unknown
  12. a cost estimate of "unknown" means unverified. never free.

Approval format: APPROVA plan_‹id›. The identifier is a placeholder, never a fake value.

About

Geneva, Switzerland

I’m Renato Baldo, an AI systems builder based in Geneva. I design agents, automations and software that run in production, with written rules and a human approving every step that matters.

More about Renato

Contact

Where should the human stay in control?

Tell me what the system should do, what evidence it can use and where it must stop.

contact@renatobaldo.com