Skip to content

Services · Track III · AI integration

Plug Midcore into what you already have.

Retrieval over your corpora. Policy-gated tool use. Evidence-chain integration into existing audit trails. The AI features your competitors are demoing — shipped, measured, certified, with a refusal surface on every consequential action.

§ A · The surfaces

Six integration patterns. Compound from day 30.

Retrieval-augmented generation

Vector + lexical retrieval over your corpora with policy-gated access. Cited answers; no hallucination.

Agent orchestration

Multi-step agents over your APIs and tools. Evidence chain on every call. Cost ceilings built in.

Document intelligence

OCR, classification, extraction, tampering detection. Powered by the Cultural Adaptation pipeline.

Voice + conversational AI

Latency-optimised voice with policy guardrails. Transcript review surfaces for analysts.

Behavioural anomaly detection

Per-vendor, per-approver fingerprints. Drift alerts with confidence. Powered by Midcore Genome.

Compliance binder

EU AI Act mapping, SOC 2 alignment, evidence packaging for the regulator.

§ B · Process

Twelve weeks. Pilot, harden, ship.

  1. Weeks 1–3

    Discovery + integration map

    Audit the surfaces where AI will land. Privacy boundaries. Data flows. Policy DSL drafted.

  2. Weeks 4–8

    Pilot integration

    Two integration patterns shipped to staging. Evidence chains wired. Refusal surfaces validated.

  3. Weeks 9–12

    Production + binder

    Pilots promoted. Compliance binder finalised. Hand-over with a runbook the on-call team can actually use.

§ C · Engagement model

Pilot → Programme → Sovereign.

Tier

Pilot

$45k–$80k

One or two surfaces wired in. Evidence chain + Charter alignment delivered. 12-week engagement.

Tier

Programme

from $28k/mo

Embedded integration team, multi-quarter horizon. Up to six surfaces, regulator-facing binder.

Tier

Sovereign

from $90k/mo

On-prem deploy patterns, air-gap-ready, dedicated senior team. For institutions with hard sovereignty constraints.

§ D · Charter alignment

The instruments behind every integration.

Every NeuroBazar engagement is anchored to specific Charter modules — each published with what it enables and what it regulates. This service leans on the instruments below.

All 20 instruments →

  • Ethics Center

    No. 03

    A policy DSL where ethical limits are code, not a poster on the wall.

  • AI Safety

    No. 04

    Runtime guardrails on every model call.

  • Vector DB

    No. 09

    Knowledge retrieval gated by policy, citation by default.

  • Evidence Room

    No. 10

    Hash-chained audit of every AI action.

  • Governance Center

    No. 12

    Policy enforcement across every tenant, model, and surface.

  • Federated Learning

    No. 16

    Models trained without centralising your data.

§ E · FAQ

Questions, frankly answered.

Will you use my data to train models?
No. Our default integration uses provider APIs configured for zero-retention; for sovereign deployments, the data never leaves your boundary. We sign that in writing on every engagement.
Which model providers do you support?
Anthropic, OpenAI, Google, Mistral, Cohere, and self-hosted open-source via the Midcore router. We treat the model as a swappable component; the wrapper is the product.
How do you handle hallucinations?
Retrieval-grounded generation by default. If a claim has no retrievable source, the system refuses to make it or labels it as opinion. Every surface ships with a "show sources" affordance the user can always reach.
EU AI Act readiness — what do you deliver?
A mapped binder covering risk classification, conformity assessment scope, the policy DSL governing tool use, evidence chains for high-risk surfaces, and a deployment register your DPO can audit.
Can you integrate with our existing LLM gateway?
Yes. We have integrated with LiteLLM, OpenRouter, internal gateways, and Crown-built sovereign deployment platforms. The Charter alignment travels with the integration; we do not require ours to be in the path.
What does "evidence chain" actually mean operationally?
Every meaningful AI action — a tool call, a retrieval, a model output — gets a hash-linked record in a chain the customer owns. The user can request their slice. The auditor can replay it. Nothing is "lost."

Next step

A scoped pilot before a long bet.