Gryffin Forge

AI automation showcase

Automation that explains itself.

Deterministic pipelines first, AI only where it earns its place — with guardrails, limits and a person in the loop. Everything below is interactive and runs on fictional records.

Workflow orchestration

Watch a record move through the pipeline.

Pick a fictional business and run the same five stages the demo engine uses: analyse, extract, route, personalise, preview.

Fictional input record

Scan results are canned for this demo; routing runs the real rules. No requests leave your browser.

  1. Website analysisqueued
  2. Structured extractionqueued
  3. Classification & routingqueued
  4. Template personalisationqueued
  5. Preview readyqueued

Template routing

Type a business name. See which template it gets — and why.

Seventeen ordered keyword rules, the same ones the production engine uses. Then open the result as a live demo.

Result

Beauty salon

Matched keyword Nail · template beauty

Display name cleaned to “Bloom Nail & Lash Bar” (legal suffix removed).

See the live beauty salon demo

rules evaluated in priority order5/17

  1. 01Pet care
  2. 02Tattoo studio
  3. 03Clinic / aesthetics
  4. 04Barbershop
  5. 05Beauty salon
  6. 06Solar & energy
  7. 07Gardens & landscaping
  8. 08Removals & transport
  9. 09Garage & car care
  10. 10Cleaning
  11. 11Trades & construction
  12. 12Home care
  13. 13Holiday stays
  14. 14Events & photography
  15. 15Fitness
  16. 16Restaurant & café
  17. 17Estate agent

Same rules as the production engine, executed in your browser. Deterministic, explainable, no AI call and nothing sent.

Why rules before a model? Read the article →

Classification & extraction

Turn a messy enquiry into a structured record.

Intent, urgency, service, postcode, budget and timing — extracted locally, with injection attempts flagged instead of obeyed.

136/800Processed locally · nothing sent

structured_output.json● rules engine

{
  "intent": "quote",
  "urgency": "normal",
  "service": "Cleaning",
  "fields": {
    "postcode": "BS6 5QT",
    "budget": "£250",
    "when": "next week",
    "size": "3 bed",
    "email": false,
    "phone": false
  },
  "confidence": 0.94,
  "flags": [],
  "next_step": "Enough detail for an illustrative estimate; confirm by a person."
}

Capabilities

What exists today, honestly labelled.

Business website analysis

In production (internal)

Multi-variant scans that detect HTTPS, mobile layout, booking platforms and photos, storing evidence with every verdict.

Lead research architecture

In production (internal)

Multiple independent signals with an explicit 'unknown' state, so the system never guesses.

Template personalisation

In production

Name, place and services applied to industry templates; legal suffixes cleaned from display names.

Workflow orchestration

In production

Staged pipelines where each step's output is validated before the next runs.

AI response classification

Prototype (rules)

Classifying enquiries by intent and urgency, with complaints always routed to a person.

Structured information extraction

Prototype (rules)

Postcodes, budgets, dates and sizes pulled from free text into a typed record.

AI-assisted content structuring

Planned — not active

A model proposes copy and template for ambiguous cases, constrained to valid outputs.

Safeguards

Built to be safe by default.

No live AI on public pages

Every demo on this page runs deterministic rules in your browser. There is no model to manipulate.

Prompt-injection aware

Instruction-like text in an enquiry is flagged and treated as data; it never changes how a record is processed.

Rate limits & size caps

Server endpoints are rate-limited, inputs are length-capped and proxies only accept allow-listed content.

Humans own the outcome

Complaints, urgent jobs and low-confidence results always go to a person.

Fictional data only

Public demonstrations never touch the internal research database or anyone's personal information.

No exposed credentials

No API keys exist in the browser bundle; future integrations are server-side and disabled by default.

Future integrations are wired, not switched on.

Interfaces exist for an external AI agent, the personalisation engine and a CMS. They stay disabled server-side until explicitly configured — and no commercial outreach is part of this release.

Architecture