How it works

How Hesper AI works

A deep dive into the detection pipeline — from document receipt to fraud verdict in under 80 milliseconds.

Where Hesper AI fits in your pipeline

Document upload

Your user / system

→

Hesper AI API

Fraud detection

→

Your OCR pipeline

Text extraction

→

Your platform

Processing / approval

Your application sends the document

Before passing a document to your OCR pipeline, send it to the Hesper AI /v1/analyze endpoint via a standard multipart POST request. We accept PDFs, PNGs, JPEGs, and TIFFs up to 50MB. No preprocessing or conversion required.

Pixel-level forensic analysis

Our computer vision models scan the document at the pixel level — detecting cloning artifacts, brightness edits, compression inconsistencies, and other traces left by editing software including Photoshop, GIMP, and AI image generators.

Font and typography analysis

We extract and analyse every character in the document, detecting inconsistent font weights, spacing anomalies, and characters that don't match the document's declared font metadata — common signs of text replacement.

Metadata and structural validation

Document metadata is cross-referenced against the visual content. Creation timestamps, software fingerprints, and embedded color profiles are checked for inconsistencies that indicate post-processing.

Verdict returned in under 80ms

A structured JSON response is returned: a fraud score from 0–100, a verdict (CLEAN / SUSPICIOUS / LIKELY_FRAUD), an array of findings with pixel coordinates, and a confidence rating. Your pipeline continues normally for clean documents, or you halt for review.

The API request and response

Request

POST /v1/analyze HTTP/1.1
Host: api.hesper.ai
Authorization: Bearer hsp_live_sk_xxxxxxxxxxxx
Content-Type: multipart/form-data; boundary=----HesperBoundary

------HesperBoundary
Content-Disposition: form-data; name="file"; filename="invoice.pdf"
Content-Type: application/pdf

[binary PDF data]
------HesperBoundary--

Response

{
  "id": "req_01HZ7KXPQR8FGBN4W3YV5",
  "fraud_score": 87,
  "verdict": "LIKELY_FRAUD",
  "confidence": 0.94,
  "latency_ms": 72,
  "document_type": "invoice",
  "page_count": 2,
  "findings": [
    {
      "type": "pixel_manipulation",
      "description": "Clone stamp artifact detected in amount field",
      "region": { "x": 142, "y": 88, "width": 168, "height": 32 },
      "severity": "high",
      "confidence": 0.97
    },
    {
      "type": "font_inconsistency",
      "description": "Character metrics deviate from document baseline",
      "region": { "x": 80, "y": 200, "width": 340, "height": 30 },
      "severity": "medium",
      "confidence": 0.88
    }
  ],
  "metadata": {
    "created_at": "2024-11-10T14:23:00Z",
    "software": "Adobe Acrobat 23.0",
    "last_modified": "2025-01-15T09:41:22Z"
  }
}

Latency and throughput specs

p50 latency< 60ms

p99 latency< 80ms

Max document size50MB

Supported formatsPDF, PNG, JPG, TIFF

Max throughput (Free)300 pages / month

Max throughput (Enterprise)Unlimited

Concurrent requestsUp to 100 (configurable)

API uptime SLA99.95% (Enterprise)

Ready to integrate?

Get API access and analyze your first document in under 5 minutes.