queued process scans split to OCR merge results review fields flag disputes query tools results verified data score fields high confidence medium low confidence
Agentic AI · System Architecture

Aviation Work Order — Intelligent Automation Pipeline --live demo !

Scanned PDFs → AI Extraction → Human Verified → Maintenance System  |  © Jinan Kordab
PHASE 1 — INGEST
PHASE 2 — PROCESS
PHASE 3 — OCR
PHASE 4 — AI AGENTS
PHASE 5 — VALIDATE
PHASE 6 — ROUTE
INPUT 📄 Scanned PDFs ~200/week
Handwritten + typed
3–5 pages each
ACTION ⚡ Queue & Schedule Email listener, shared folder
watch, web upload
→ Redis Streams
ACTION 📑 Document Processing
🖨️Convert each page to high-res image
🔧Straighten, clean & improve scan quality
🖼️Describe embedded photos with AI
🟠 MATLAB — Image Processing
ACTION 👁️ Text Recognition Reads handwritten
and typed text
ACTION 📐 Location Mapping Locates each word
on the page
ACTION 🔤 Character Correction Fixes misread letters
and numbers
🟠 MATLAB — Deep Learning CNN
AGENT 🤖 Extractor Reads text, fills in
all structured fields
AGENT 🔍 Critic Checks every field
for errors
AGENT ⚖️ Reconciler Fixes only the
disputed fields
ACTION — REAL-TIME LOOKUP 🔌 MCP Tool Servers
✈️Check aircraft registration is valid
🔩Verify part numbers in catalog
📋Confirm safety directives are current
ACTION — RULE-BASED ✅ Deterministic Validation
🔢Tail number format check
🔩Parts exist in catalog
📋Compliance cross-check
🔄Duplicate detection
🟠 MATLAB — Confidence Calibration
DECISION 🎯 Confidence
Router
How confident is
the AI in each field?
OUTPUT 🟢 Auto-Approve Push directly
to maintenance system
OUTPUT 🟡 Human Review Clerk checks
flagged fields
OUTPUT 🔴 Manual Entry Clerk types
from scratch
🔵 AutoGen — Multi-Agent AI (Phase 4)
3 AI agents work as a team to cross-verify each extraction:
Extractor — reads all OCR text and fills in structured fields (aircraft ID, parts, hours).
Critic — reviews every field against the original scan and queries MCP tools to verify against real databases.
Reconciler — re-processes only flagged fields, producing a final high-confidence result.
Catches AI hallucinations before they reach validation. Adds ~15–25s latency, boosts accuracy by 2–4%.
🟠 MATLAB — Signal & Image Processing (Phases 2–3 & 5)
4 MATLAB toolboxes run behind the scenes across multiple phases:
Image Processing — deskew, denoise & enhance scans before OCR sees them.
Deep Learning — CNN classifier fixes aviation-specific character confusion (0/O, 1/l, B/8, 5/S).
Statistics — calibrates raw AI confidence scores into true probabilities (Platt scaling).
Predictive Maintenance — Weibull failure analysis & parts demand forecasting from work order history.
Runs via MATLAB Engine API for Python inside the pipeline. ~$50/wk license.
🟢 MCP — Model Context Protocol (Phase 4)
A standard tool interface that gives AI agents access to real company databases during extraction, not after.

Where: Plugs into the Critic agent — called in real-time while checking fields.
What: Validates aircraft IDs against Fleet Registry, parts against catalog, safety directives against compliance database.
Why: Catches wrong data one step earlier. Model-agnostic — works with any LLM, no vendor lock-in.
Action
Decision
Info / Input
Ingest
Process
OCR
AutoGen
MCP
MATLAB
Validate
Route