Problem

Medical job listings are spread across many Australian portals

Solution

Source discovery and manual-link ingestion collect job opportunities into one platform.

Problem

Job pages change layout and break scrapers

Solution

Adaptive extraction pipeline using site detection, adapters, metadata extraction, semantic fallback, AI-assisted recovery, validation, confidence scoring, and review queue.

Problem

Fixed CSS selectors are fragile

Solution

Semantic extraction looks for field patterns such as Position Title, Job Title, Classification, Salary, Location, and Closing Date.

Problem

Some pages include hidden structured metadata

Solution

Structured metadata extraction reads JSON-LD, embedded metadata, structured page data, and application state objects.

Problem

Low-confidence extraction can pollute decisions

Solution

Confidence scoring routes high-confidence records to auto-approval, medium-confidence records to repair, and low-confidence records to human review.

Problem

Recruitment sites may block automated fetching

Solution

Blocked or robots-related jobs are moved into manual review with safe display fallbacks.

Problem

Candidate suitability is not obvious from a job title alone

Solution

Candidate-aware scoring evaluates direct fit, eligibility blockers, hidden opportunities, red flags, and recommended action.

Problem

AMC/IMG candidates need pathway-sensitive filtering

Solution

Analyzer detects AHPRA, AMC, IMG, supervision, WBA, hospital, rural, and shortage-area signals.

Problem

Manual review needs structure

Solution

Review queue stores reason, severity, status, source, adapter, confidence, missing fields, red flags, assignment, and reviewer notes.

Problem

Extraction failures need monitoring

Solution

Extraction logs, diagnostics, drift alerts, and golden tests track reliability.

Problem

Workflows should not rely only on synchronous requests

Solution

Execution tasks support status, priority, payloads, retries, leases, idempotency keys, dead-letter state, and worker tracking.

Problem

Admins need operational visibility

Solution

Frontend includes Dashboard, Analytics, Drift Alerts, Extraction Health, Extraction Queue, Golden Tests, Jobs, Review Queue, Workflows, Notifications, Candidate, and Settings.

React 19, Vite, TypeScript, Tailwind CSS, React Router, Lucide icons. Authenticated operations console with protected routes.

FastAPI app with versioned API routes under /api/v1. Routers for auth, jobs, candidate profiles, candidates, discovery, scheduler, scheduled jobs, workflows, workers, review queue, operations, notifications, dead-letter, execution tasks, sources, and system APIs.

SQLAlchemy with PostgreSQL. JSONB used across jobs, candidates, diagnostics, tasks, notifications, and scheduled jobs.

Deterministic adapters, structured metadata extraction, semantic fallback, AI-assisted recovery, validation, confidence scoring, drift detection, human review.

Scheduled jobs, workflow runs, execution tasks, and workers. Execution tasks support retries, max attempts, leases, idempotency keys, replay counts, dead-letter information, and worker ownership.

Workers track name, type, hostname, process ID, version, status, capabilities, current task, completed count, failed count, start time, heartbeat, and stopped time.

Application folder creation per job.

JWT access tokens signed with HS256, PBKDF2-SHA256 password hashing, email normalization, token expiry, bearer authentication, and role checks.

Local-first development behavior. Backend CORS allows Vite dev origins. FastAPI backend is local-only.

Australian job sites and recruitment engines through allowed hosts and source toggles, including WA, QLD, NSW, NT, SA, VIC, TAS health sources, SEEK, Indeed, PageUp, Workday, Taleo, and hospital career portals.

User Login

Operations Console

Manual Job Link or Discovery Run

URL Validation and Normalization

Source Detection

Adapter Selection

HTTP Fetch

Browser Render Fallback if enabled

Structured Metadata Extraction

Semantic Field Extraction

Parser and Reliability Validation

Confidence Scoring

Candidate-Aware Analysis

Match Score and Recommended Action

Save Job and Diagnostics

High confidence becomes processed job, low confidence/blocked/mismatch goes to Review Queue

Reviewer Correction if needed

Reanalysis

Candidate Summary/Fit Analysis/Resume Generation

Notifications and Workflow History

Challenge

External job sites change HTML/CSS

Solution

Multi-layer extraction pipeline

Challenge

Scrapers can return wrong/contaminated data

Solution

Semantic mismatch checks, contamination detection, confidence caps, manual review

Challenge

SPA-rendered pages

Solution

Playwright/browser fallback

Challenge

Protected structures

Solution

PageUp/Taleo/Workday/MedCareersWA safeguards

Challenge

Robots or blocked fetching

Solution

Manual-review status and safe fallbacks

Challenge

Duplicate processing

Solution

URL uniqueness and idempotency keys

Challenge

Long-running work blocking API

Solution

Queue-compatible tasks, workers, leases, retries, dead-letter

Challenge

Candidate fit complexity

Solution

Direct fit, blockers, hidden opportunity, red flags, recommended action

Challenge

Review traceability

Solution

Review queue metadata

Challenge

Failure detection

Solution

Drift alerts, logs, diagnostics, golden tests

Challenge

Sensitive logs

Solution

Redaction and safe errors

Challenge

Unauthorized access

Solution

JWT and RBAC