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.
-
Weeks 1–3
Discovery + integration map
Audit the surfaces where AI will land. Privacy boundaries. Data flows. Policy DSL drafted.
-
Weeks 4–8
Pilot integration
Two integration patterns shipped to staging. Evidence chains wired. Refusal surfaces validated.
-
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.
-
⚖ 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."