AI-Powered Middleware

AI Answers
That Make Sense

Centralize · Validate · Convert

The control layer between your business and AI. Define prompt endpoints, enforce validation rules, map outputs to your data model — and ship reliable AI features without the chaos.

Requestador — the AI control layer
integration
validators
0 deploys for rule changes
What is Requestador?

Your AI
Control Layer

Requestador sits as intelligent middleware between your applications and AI models. Instead of wrestling with unpredictable outputs, you define structured prompt endpoints with built-in validation, business logic, and automatic retry mechanisms.

Think of it as an API gateway built specifically for AI interactions — complete with parameters, validation rules, and guaranteed response formats. When AI responses don't match your requirements, Requestador automatically retries with refined prompts or applies transformation rules.

Whether you're enriching PIM data, validating banking transactions, generating product imagery, or powering multi-tenant SaaS — every AI output is accurate, structured, and production-ready from day one.

ENDPOINT DEFINITION
POST /api/prompts/extract-product-data { "prompt_template": "Extract from: {input_text}", "parameters": { "input_text": "required", "format": "json" }, "validation_rules": [ "price_must_be_numeric", "category_in_catalog" ], "retry_on_failure": true, "max_retries": 3 } // Validation + auto-retry on failure if (invalid) → refine_prompt() → retry()
Platform Features

Infrastructure Built for
Production AI

Every feature designed to eliminate AI unpredictability and deliver structured, business-ready responses your systems can rely on.

Centralized Prompt Hub

Define all AI interactions like API endpoints. Set parameters, expected outputs, and validation rules in one place. Version control, team sharing, zero-code deployments.

Smart Validation Engine

Custom validation logic runs before responses reach your app. Check types, verify values, enforce business rules. Auto-retry on failure with progressively refined prompts.

Catalog Mapping

Map AI's natural language outputs to your canonical values. Define catalogs for products, categories, codes. "0.5L" or "500ml" becomes "Half-Litre Bottle" automatically.

Auto-Retry & Refinement

Responses that fail validation automatically retry with refined prompts. Configure strategies, set max attempts, define fallbacks. Never expose raw AI errors to your systems.

Multi-Tenant Access Control

Organization-aware architecture built in. Define endpoints and catalogs per tenant with isolated data and prompts. Perfect for SaaS platforms with varying client requirements.

Private AI Compatible

Integrates with on-premise private AI deployments. Sensitive data never leaves your infrastructure. Full auditability and data residency compliance for regulated industries.

How It Works

Three Steps to
Reliable AI

Requestador transforms unreliable AI outputs into production-ready structured data your business can trust at scale.

01

Define Prompts

Create API-like endpoints with parameters, validation rules, and output formats. Set up business logic, catalog mappings, and retry strategies — all in one place without developer involvement for every change.

02

Validate & Transform

Every AI response runs through your validation rules before reaching your application. Check types, verify against catalogs, apply business logic. Failures trigger automatic retries with refined prompts.

03

Receive Clean Data

Get validated, business-ready data in your required format — mapped to your domain model, typed correctly, and guaranteed to match your specifications every time.

Industry Solutions

Requestador
in the Wild

Five real teams. Five different problems. One control layer that made AI reliable in production.

E-Commerce

One Integration,
Unlimited Validators

A fast-growing e-commerce operation was scaling into new markets, adding carriers, and integrating more systems every quarter. But validation logic — the rules that check whether an address is real, a VAT number is valid, a shipment weight is within limits — was being rebuilt from scratch every single time, in every single system.

ARCHITECTURE
// Before: validation everywhere
Storefront  → own rules
Backoffice  → own rules
Warehouse   → own rules
Call center → own rules

// After: integrate once
Storefront  → Requestador → ERP
Backoffice  → Requestador → WMS
CSV Import  → Requestador → Orders
Call center → Requestador → CRM

// Add new validators — zero deploys
01
The Problem

Validation logic lived everywhere — and it was breaking everything

When a new carrier required a different address format, four teams had to coordinate a release. When a new market launched, developers rebuilt the same postal code checks they'd written twice already. Customers were getting through checkout with bad addresses and finding out at the warehouse. VAT numbers were passing the storefront but failing invoicing. The same broken data, over and over, from different angles.

Duplicate rules across storefront, ERP, warehouse, and backoffice — each slightly different

Address errors slipping through to the warehouse, causing failed deliveries and support backlogs

Carrier rule changes forced coordinated multi-system deployments across teams

Free-text return reasons made analytics meaningless and routing entirely manual

02
The Solution

Requestador became the single place where validation lives

Instead of rebuilding rules per system, the team integrated each application once to Requestador. From that point on, adding a new validator — a new carrier's weight limit, a new country's address format — meant configuring a single endpoint. No deployment. No cross-team coordination. Business teams started owning rules that used to require a developer ticket.

How it works

1

Checkout submits address to the validate_address endpoint. Requestador checks postal code against city, validates phone format, confirms carrier compatibility — and returns structured errors or a normalized result.

2

Same endpoint reused in backoffice order editing, call center entry, CSV bulk imports. The rules are central — every surface gets identical behavior without any extra code.

3

Return reasons are mapped through a catalog layer: "product didn't fit" → SIZE_ISSUE, "arrived damaged" → DAMAGE_TRANSIT. Analytics and routing automation become possible overnight.

Any System Requestador Validated Result Continue Process
03
The Result

Validation stopped being a development task

The team stopped rebuilding the same logic. New validators were added in minutes by business teams, not developers. Delivery failures dropped. Support workload dropped. And when a new market launched, integration was a configuration task — not a sprint.

Integration point for all validation
New validators without new code
0Deployments for rule changes
Delivery failures & support cost
Product Data · Pimcore

50,000 Products.
Hours, Not Months.

An e-commerce operator running Pimcore had a catalog problem. Tens of thousands of products — imported from suppliers, added over years — with inconsistent names, missing descriptions, and zero SEO structure. Manual enrichment was happening but it was never going to catch up. They needed AI. But they needed AI they could control.

ENRICHMENT PIPELINE
// Step 1: Classify product
"endpoint": "classify_group"
"output":   { "group": "DIY" }

// Step 2: Generate group-specific content
"endpoint": "generate_diy_content"
"output": {
  "product_name": "Power Drill 18V Compact",
  "short_desc":   "Lightweight cordless...",
  "seo_title":    "Buy 18V Compact Drill..."
}
// Catalog validates → writes to Pimcore
01
The Problem

Content creation was a bottleneck that only got worse at scale

The catalog had grown faster than the team could enrich it. DIY products needed technical descriptions. Wine needed storytelling. Furniture needed dimensions and use-case language. Each category had its own logic — and every time that logic needed to change, a developer had to be involved. AI was an obvious answer. But every AI experiment produced outputs that were inconsistent, hard to validate, and difficult to feed back into Pimcore in a structured way.

Thousands of products with incomplete or inconsistent names and descriptions

Different product groups needed completely different content logic — impossible to generalize

Every prompt or logic change required a developer and a deployment cycle

AI outputs weren't structured enough to write back directly into Pimcore fields

02
The Solution

A Pimcore plugin turned AI into a controllable, configurable content engine

The team built a Pimcore plugin that connected directly to Requestador endpoints. Each action in Pimcore mapped to an endpoint that defined the prompt logic, expected output structure, and validation rules. When a product manager selected 500 products in the datagrid and clicked "Enrich", Requestador handled classification, prompt selection, AI generation, validation, and field mapping — automatically.

How it works

1

The classify_group endpoint determines what kind of product it is — DIY, Furniture, Wine, Electronics — based on raw input data. This classification drives everything that follows.

2

A group-specific content endpoint generates a product name, short description, long description, and SEO content using prompts tailored to that category. Wine gets sensory language. DIY gets specs and compatibility.

3

The catalog validation layer normalizes the AI output — fixing casing, enforcing character limits, checking against allowed value lists — before writing structured data back into Pimcore fields.

Pimcore Grid Requestador Private AI Catalog Validation Pimcore Fields
03
The Result

AI became an operational tool, not an experiment

What previously took weeks of manual copywriting was reduced to hours of automated processing. More importantly: the business team could now adjust prompt logic themselves. No tickets. No deployments. No waiting. AI was finally under control.

Hoursto enrich thousands of products
0Dev deploys for prompt changes
100%Business-owned prompt logic
SEOValidated output on every product
Banking & Finance

Six Systems.
One Source of Truth.

A mid-sized European bank had built its digital infrastructure piece by piece over a decade. Six systems. Six sets of validation rules. When a customer opened an account on mobile, the data looked valid. When it reached the compliance system three steps later — it didn't. The bank needed a single validation layer. And it needed AI — but couldn't send a single byte of customer data outside its own walls.

ONBOARDING VALIDATION
// Private AI normalizes input
"raw_address": "Main str. 12a, Zagreb"
"normalized":  "Main Street 12A, Zagreb"

// Rule-based final decision
"status":          "pass"
"blocking_errors": []
"audit_metadata": {
  "rule_version":  "v4.2.1",
  "ai_confidence": 0.97,
  "data_residency": "on-premise"
}
01
The Problem

Validation inconsistency was creating compliance risk at every touchpoint

Each system had its own interpretation of what valid customer data looked like. A mobile onboarding form would accept an abbreviated street name. The core banking system would reject it two steps later, triggering a manual review queue that was already overwhelmed. When regulations changed, the compliance team had to coordinate simultaneous releases across six teams — or accept a window of inconsistency that regulators could flag. And when the bank wanted to explore AI for smarter data normalization, GDPR and banking supervision rules made public AI services completely off the table.

Data accepted by mobile banking rejected later by compliance — creating manual review backlogs

Regulatory updates required coordinated releases across six independent systems

GDPR and banking supervision blocked any use of public AI services with customer data

No audit trail for validation decisions — impossible to explain outcomes to regulators

02
The Solution

Requestador became the gateway — with a private AI model running entirely on-premise

All digital channels were routed through Requestador before data reached core banking or compliance systems. A private AI model, deployed inside the bank's own infrastructure, handled the interpretive work — normalizing addresses, extracting data from uploaded documents, flagging ambiguous inputs. Requestador then applied deterministic rule-based validation on top of the AI output, ensuring every decision was explainable, versioned, and auditable.

How it works

1

Customer submits onboarding data via any channel. All channels route to Requestador — no system talks directly to core banking anymore.

2

Where interpretation is needed — normalizing an address, reading a document — the private AI model processes the data on-premise. No customer data leaves the bank's infrastructure.

3

Requestador applies rule-based final validation on the AI output. The result includes a versioned rule ID and confidence score — a full audit trail that satisfies regulatory requirements.

All Channels Requestador Private AI (on-prem) Core Banking
03
The Result

The bank got AI and kept control — with zero data leaving its infrastructure

Onboarding inconsistencies dropped. The manual review queue shrank. Regulatory updates became single-system changes. And for the first time, the bank had a complete audit trail for every validation decision — with the confidence to demonstrate compliance to regulators.

1Central place for all regulatory rules
0Customer bytes sent to public AI
Manual review workload
FullAudit trail on every decision
AI Image Generation

From Design Bottleneck
to Automated Pipeline

A large e-commerce company launched product bundles constantly — starter kits, seasonal promotions, B2B clearance sets. Each bundle needed a custom visual: the right products composed together, the right background for the campaign type, the right format for each channel. Design was always the last step and always the slowest one.

IMAGE PIPELINE
// Category-driven layout rules
"diy_products""neutral technical bg"
"lifestyle_bundle""contextual scene bg"
"promo_bundle""badge space reserved"

// Output validation before publish
"resolution":   "1200x1200"  
"aspect_ratio": "1:1"        
"brand_rules":  "compliant"  
// → Auto-stored DAM, linked to products
01
The Problem

Marketing moved fast. Design couldn't keep up.

The commerce team could create a bundle in five minutes. Getting an image for it took hours — sometimes days. Designers were handling production assets for short-lived promotional bundles, work that was repetitive but couldn't be skipped. Visual inconsistency crept in across campaigns. And as the number of bundles grew, the design dependency became a hard ceiling on marketing velocity. Public AI image services were tempting — but product images couldn't leave company infrastructure due to IP and branding concerns.

Hours to days of designer time per bundle — campaigns delayed waiting for visuals

Inconsistent layouts, shadows, and positioning across different campaigns and designers

IP and branding concerns made public AI image services unusable

No way to scale visual production alongside growing bundle volume

02
The Solution

Requestador turned bundle creation into an automatic visual pipeline

Requestador was deployed as the orchestration layer between the commerce system and a private AI image generation service running inside company infrastructure. When a bundle was created, Requestador received the product composition data and automatically determined the correct layout, background style, and brand rules based on product category and campaign type. The image was generated, validated against technical requirements, and published to the DAM — without a designer ever touching it.

How it works

1

A bundle is created in the commerce system. Requestador receives the product list and determines layout rules: DIY products get neutral technical backgrounds; lifestyle bundles get contextual scenes; promo bundles reserve space for a discount badge.

2

The private AI image service generates a composed bundle image on-premise — consistent lighting, standardized spacing, brand-compliant backgrounds. Product images never leave company infrastructure.

3

Requestador validates the output — resolution, aspect ratio, file naming — before returning it. Non-compliant images are regenerated automatically. Approved images are stored in DAM and linked to the bundle.

Bundle Created Requestador Private AI Validated DAM / Published
03
The Result

Image production became part of the operational workflow

Bundle image creation dropped from hours to minutes. Campaign launches accelerated. Visual consistency improved — not because designers were more careful, but because the rules were encoded. Design capacity was freed for creative work that actually needed a human. And bundle experimentation increased because the cost of producing a short-lived promotional asset became nearly zero.

Minvs hours per bundle visual
100%Brand-consistent outputs
0Customer IP sent to public AI
Bundle experimentation velocity
Enterprise SaaS

Multi-Tenant Complexity.
Clean Product Core.

An enterprise SaaS platform had a problem that every successful SaaS eventually faces: too many customers, each with slightly different requirements. One client needed a mandatory field others didn't. Another had country-specific validation rules. A third integrated an ERP that used status codes nobody else had ever seen. The platform was accumulating conditional logic, tenant-specific branches, and custom integrations — and it was getting harder to ship anything.

MULTI-TENANT VALIDATION
// Tenant-specific endpoint
"org_id":   "tenant_42"
"endpoint": "validate_import_record"

// External system mapping
"external_status": "DELIV_OK""internal_status": "Completed"

// Config safety check
"check_dependencies": true
"detect_conflicts":   true
"safe_to_activate":   true
01
The Problem

Customization was accumulating as technical debt inside the product

The platform had grown through three approaches to tenant variation: hardcoded exceptions, configuration flags scattered through the codebase, and custom integration branches for large clients. All three created the same outcome — a product that was increasingly difficult to change without breaking something. Validation logic was particularly fragile: it touched every data entry point, import pipeline, and external integration, and every tenant had a slightly different version of what "valid" meant. Business rules changed faster than releases could ship them.

Tenant-specific validation hardcoded as conditional branches — the product was becoming unmaintainable

Import pipelines failing on inconsistent external formats — support costs growing with every new integration

ERP and legacy system mappings built custom per client — no reusable pattern

Business rules changing faster than release cycles — customers waiting weeks for simple rule updates

02
The Solution

Validation moved out of the product and into a dedicated configurable layer

Requestador was introduced as the validation and decision layer between client input and the SaaS platform's business logic. Instead of encoding tenant-specific rules in code, each tenant got its own Requestador endpoints. The platform stayed stable — it called Requestador and received a consistent, structured response. The rules lived outside the product, which meant they could be changed by business teams without touching the codebase, and new tenants could be onboarded with their own configuration rather than a new development sprint.

How it works

1

Client data enters the SaaS platform and is routed to tenant-specific Requestador endpoints. Tenant 42's rules are different from Tenant 7's — but the platform always calls the same interface and gets the same response structure.

2

Imports and external integrations pass through Requestador's mapping layer. ERP status code "DELIV_OK" becomes the platform's internal "Completed" — without custom code per client.

3

Configuration changes are validated before activation — Requestador checks dependencies and conflict risks, preventing the runtime failures that previously required emergency hotfixes.

Client Input SaaS Platform Requestador (per-tenant) Business Logic
03
The Result

The product got cleaner while the platform got more flexible

Tenant-specific branches disappeared from the codebase. Release cycles shortened. Support costs from bad imports dropped. Business teams started managing validation rules themselves — the work that used to require a developer ticket now happened in a configuration interface. New enterprise clients could be onboarded with custom rules without any development work.

Tenant-specific code branches
0Deployments for rule changes
Import failures & support cost
CleanProduct core maintained
Who It's For

Built for
AI-Forward Teams

SaaS Companies

E-commerce Teams

Product Managers

Banking & Finance

Developers

AI Engineers

Ready for Reliable AI?

Join teams powering production AI with Requestador.