Reference architecture

ARCHITECTURE

How we build agentic systems that act on real business data — one orchestrator, specialised engines, explicit human gates and a durable execution layer underneath.

  • 3 planesExperience, control and data — separated by contract
  • Gate-firstNothing costly or external runs without approval
  • Vendor-neutralModels and providers swap without rewriting workflows
Auto-tour the system Talk to us

Operating model

One brain, two coordinated engines and one measurable execution loop.

  1. You set the goal

    “We are launching product X.”

  2. Orchestrator

    One assistant, one workspace — plans, routes and explains every step.

  3. Shared business memory

    Brand rules · products · customers · performance · approved assets

  4. Engine A

    Decision engine

    1. Diagnose
    2. Prioritise
    3. Test
    4. Improve
    Engine B

    Production engine

    1. Research
    2. Concept
    3. References
    4. Prototype
    5. Variants
    6. Channel assets

  5. One task graph

    Clear inputs, outputs, owners and completion criteria

  6. Your approval

    Direction, assets, spend and release scope

  7. Execution across connected tools

    Generate · prepare · publish · update · measure

  8. Measurement & learning

    What performed · what changes next

Results feed the next decision and the next creative iteration

Capability map

Work is routed by task through a single entry point, then handled by the engine and skills that own it.

The assistant is the control surface, not the storage layer. Production assets, secrets, durable jobs and provider integrations live in the backend.

System architecture

Three planes with explicit contracts between them. Workflows are versioned graphs, every stage is durable, and vendors sit behind adapters.

Experience

Where people work

  • Client

    Sets goals, reviews results, approves scope

  • Private assistant

    One entry point, skills and task contracts

  • Language model

    Reasoning, planning and orchestration

Tool calls and context Execution evidence

Control plane

Where the rules live

  • MCP / tool gateway

    Identity, permissions and domain tools

  • Application service

    Business rules, context assembly, validation

  • Workflow registry

    Versioned task and production graphs

  • Approval gate

    Scope, asset version, cost and release permission

  • Durable job graph

    Idempotent, resumable stages

  • Workers

    Deterministic execution

  • Provider adapters

    Swap models and vendors without changing the contract

Operations, async jobs, approved releases Results and performance signals

Data & providers

Where state and effects live

  • Relational store

    Tasks, catalogue, approvals, versions, lineage, audit log

  • Object storage

    Sources, references, previews and final assets

  • Business data

    Commerce, analytics, advertising and CRM

  • Generation providers

    Image, video and voice

  • Delivery channels

    Draft or release — only after approval

Staged execution

Cheap and reversible first, expensive and external last — with a human gate between them.

  1. Context pack

    Product facts, prior results, approved references and current assets.

  2. Versioned plan

    Stages, deliverables and an estimated cost.

  3. Human gate

    Approve direction and spend

    The exact plan version and the cost limit are recorded.

  4. Stage 1

    Prototype

    Research, prompts and low-cost previews. Outputs, lineage and cost are stored.

  5. Stage 1

    Review

    Regenerate, adjust or choose a direction.

  6. Human gate

    Approve production

    Only the selected direction moves on.

  7. Stage 2

    Production

    Asynchronous jobs for video, variants and channel assets. Provider receipts are kept.

  8. Stage 2

    Ready for release

    Final assets with versions and evidence.

  9. Human gate

    Approve release

    Publishing or a campaign action needs explicit permission.

  10. Release

    Publish or deliver, with the external receipt or campaign ID.

  11. Outcome

    What shipped, what it cost, what performed and what to iterate next.

Job lifecycle

Production is a set of resumable states with explicit review and reconciliation — never one opaque “run” call.

Main path

  1. DRAFT Brief and required context complete
  2. VALIDATED Run cheap, reversible stages
  3. PROTOTYPING Previews stored and QA-checked
  4. REVIEW_REQUIRED Direction selected
  5. APPROVED_FOR_PRODUCTION Exact versions and cost limit recorded
  6. PRODUCTION_QUEUED Worker lease acquired
  7. PRODUCTION_RUNNING Outputs confirmed and QA passed
  8. READY_FOR_RELEASE Authorised external action confirmed
  9. RELEASED Performance window available
  10. MEASURED Next iteration or a new hypothesis

Review, failure and reconciliation

  • REVIEW_REQUIRED→PROTOTYPINGRegenerate or adjust
  • PRODUCTION_RUNNING→FAILEDConfirmed failure
  • PRODUCTION_RUNNING→PENDING_RECONCILIATIONProvider-side effect unknown
  • PENDING_RECONCILIATION→READY_FOR_RELEASEResult confirmed
  • PENDING_RECONCILIATION→FAILEDNo result confirmed
  • READY_FOR_RELEASE→ARCHIVEDAsset kept without release

Production pipeline

A staged graph from brief to performance signal. Winners get multiplied — not everything.

  1. Brief

    Product · audience · objective · approved claims

  2. Research

    Current facts and asset inventory

  3. Concept & script

    Hook · story · offer angle

  4. Visual references

    Style rules and scene constraints

  5. Prototype stills

    Low-cost generation first

  6. Review gate

    Quality · brand fit · product fidelity

  7. Enhancement

    Upscale · retouch · product placement

  8. Motion & voice

    Video · lip sync · voice

  9. Controlled variants

    Audience · format · script

  10. Channel package

    Social · ads · CRM · website

  11. Performance signals

    Fatigue · conversion · winners

Design principles

The rules every system we deliver is built on.

The assistant is the control surface — not the storage layer.

Core rule

Have a system to build?

Product intake

Talk to us

Security: security@neuropathicsystem.com