AI Agents
Discovery Express

Platform · Layer 04

Evals from day one, audit on every action.

An agent you cannot measure is a liability with a nice interface. Evals gate every release, drift is watched, every action leaves a trail.

Talk to an architectBack to the platform
An aluminum shield behind a thin frosted glass plate
04 / 05

Components

What runs in this layer.

Evals · run 312
Demo data
  • Resolution accuracy
    96%
  • Policy compliance
    99%
  • Tone and language
    94%
  • PII masking
    100%
  • Escalation timing
    91%
  • Cost per run
    97%

Commitment gate

KPI 60% agreed · 63.8% measured · no drift in 14 days

DaCodes Evals · LangSmith · weekly quality report

07 components in this layer

  • In production× 6
  • Reference architecture× 1
01

DaCodes Evals framework

No passing run, no pilot. Written with your team.

02

LangSmith and Braintrust

Tracing in your account, datasets in your repo.

03

Guardrails and PII masking

Topic limits and redaction before a model sees anything.

04

Audit trail

Every tool call, decision, approval and model version.

05

Cost controls

Token limits, monthly budgets and alerts at your threshold.

06

Drift detection

Scheduled runs on production samples. A drop pages a person.

07Reference architecture

ISO 42001 mapping

Controls mapped. ISO 27001 is the one held today.

In production: A real client, or DaCodes itself labeled as internal use, with a KPI measured in the Proof Ledger.

Operating controls

Ten controls, on by default.

None is an add-on. They ship with the first pilot and stay on.

Demo data. Agent Console v0.1, Sep 2026.

Evals loop

Define, run, gate, monitor, retrain.

The loop that decides whether a change ships and whether an agent stays in production.

Evals loop · support agent v14
Demo data
  1. 01

    Define

    eval set
    v14 · 312 cases
    threshold
    ≥ 95.0%

    Eval set and thresholds agreed in discovery.

  2. 02

    Run

    cases scored312 / 312

    Every prompt, tool or model change runs the set.

  3. 03

    Gate

    96.2%Pass · ships

    vs 95.0% threshold

    v13 · 93.4% · blocked

    Below the threshold, nothing ships.

  4. 05

    Retrain

    +18 cases

    from the week 6 failures

    set 312 → 330

    Failures become new cases and prompt fixes.

  5. 04

    Monitor

    production samples, weekly

    week 6 · drift · ops lead paged

    Production samples scored on schedule; drift pages a person.

Back to define → 01

DaCodes Evals · LangSmith · Braintrust

Regulatory mapping

Built to be audited.

What we hold, what we have mapped and what is on the roadmap.

An information-security review over a printed network architecture diagram in a glass meeting room in Mexico City

ISO/IEC 27001:2022

Held

Information security, certified December 2025.

NIST AI RMF

Mapped

Mapped to the controls above.

ISO/IEC 42001

Roadmap

Controls mapped, certification not yet held.

LFPDPPP

México

Consent, ARCO rights and privacy notices in every channel.

GDPR

When applicable

Lawful basis, minimization and EU residency.

Certificates and mapping documents available under NDA. Reviewed Sep 2026.

Questions

Architects ask.

Your team and ours, together, during discovery. Your people know what a correct answer looks like; we turn that into cases, scoring rules and thresholds. The set lives in your repo and grows with every incident.

The scheduled run flags the drop, a person on our pod and one on yours get paged, and the agent falls back to the conservative path you defined (usually: escalate more). The fix ships after the eval set passes again.

ISO/IEC 27001:2022, certified in December 2025, and AWS Advanced Tier Services Partner. NIST AI RMF is mapped; ISO 42001 is on the roadmap. We list only what we can show you the document for.

Talk to an architect

45 minutes with the people who run the platform. No slides.

We reply in under one business day.

Loading the form