VIZION AI
Public Beta • Custom GPT + Edge Gateway

Czechia Expat Job Scout

Finding English-speaking work in Prague or Brno shouldn’t mean checking twenty bookmarked career pages every morning. The Czechia Expat Job Scout connects ChatGPT to an audited registry of Czech employers, direct company applicant tracking feeds, and niche boards—without storing your CV or personal data.

LIVE DIRECTORY

2026 Audited Registry

v0.4.1
  • 50+ Curated Czech Sources: Direct company applicant tracking boards (Greenhouse, Lever, Ashby, Workday), startup directories, and expat tech communities.
  • Zero CV / Personal Data Storage: Edge gateway strips contact details, phone numbers, and raw CV text before queries reach the database.
  • Two Bounded Semantic Verbs: The assistant can only call searchApprovedSources and submitSourceDiscovery.
  • Two-Tier Proposal Staging: User-suggested job boards stage in Discovery_Queue for human curator review.
DIRECT CHANNELS

Audited Company ATS Endpoints Over Recruiter Reposts

Most job aggregator listings in Central Europe are third-party agency reposts, stale duplicates, or generic lead traps. The Job Scout indexes direct corporate applicant tracking feeds (Greenhouse, Lever, Ashby, Workday) and verified Czech employers so candidates apply directly to actual hiring teams.

DATA PRIVACY

Zero-Retention Edge Validation (GDPR Compliant)

We do not store candidate resumes, phone numbers, or personal emails. An edge gateway scans all incoming requests and actively rejects personal contact data before queries reach the database. Only anonymized search parameters pass through.

REGISTRY INTEGRITY

Sandboxed Discovery Proposals Over Direct Writes

When a user suggests a newly found company hiring board, the assistant cannot modify the master directory. The proposal is appended to an isolated discovery queue with an audit timestamp, waiting for manual human curator review of domain identity and hiring legitimacy.

WHY THIS EXISTS

Why conversational AI needs deterministic guardrails before touching live data.

Most attempts to connect ChatGPT or conversational assistants to internal business data fail in one of two ways.

Either the assistant is given static uploaded files that go stale within weeks, or it is granted unmediated API access to a live spreadsheet where a single prompt injection, network timeout, or hallucination can corrupt records or leak private data.

We built the Czechia Expat Job Scout as an open proof-of-work project to solve both sides of this problem for the Prague and Brno job markets.

Job seekers get a fast, conversational search tool across 50+ curated Czech hiring channels, direct company applicant tracking portals, and verified English-speaking roles. But behind the chat window, the integration runs on strict zero-trust engineering: a Cloudflare Worker edge gateway, semantic OpenAPI contracts, automated personal data scanning, and human-in-the-loop review queues.

CURRENT SYSTEM SNAPSHOT

Live Search & Verified Data

  • 50+ Audited Czech Sources: Direct employer applicant tracking feeds (Greenhouse, Lever, Ashby, Workday), startup lists, and verified expat boards.
  • Zero CV or Personal Data Retention: Edge worker actively rejects personal contact data before database access (code: prohibited_personal_data).
  • Two Bounded Semantic Operations: Assistant is limited strictly to searchApprovedSources and submitSourceDiscovery.
  • Two-Tier Proposal Staging: Community suggestions append to Discovery_Queue with matching Change_Log audit entries.
  • Audited GDPR Compliance: Public data protection notice active at /privacy/gpt-job-search.

Need to connect a conversational assistant or customer search tool to verified company records without exposing your private database?

LET'S DISCUSS YOUR DATA WORKFLOW OVER COFFEE →

UNDER THE HOOD

Six invariant mechanisms that keep search safe and accurate.

The Job Scout is not a wrapper or a raw scraper. It is a governed, serverless architecture that enforces strict data boundaries between OpenAI’s chat interface and our internal records.

Curated Data

50+ Source Categorized Registry

Maintains an audited directory of Czech employers, tech hubs, and applicant tracking endpoints categorized by role family, English viability, and remote flexibility.

Privacy Boundary

Edge Personal Data Scanner

A Cloudflare Worker inspects every incoming request, actively discarding names, phone numbers, postal addresses, and pasted CV text before querying the database.

API Boundary

Semantic Action Contract

Enforces OpenAPI 3.1 specifications restricting the assistant to named business operations, preventing generic database CRUD or cell updates.

State Integrity

Idempotency & Replay Handling

Cryptographic idempotency keys ensure network timeouts and mobile retries return confirmed receipts without generating duplicate queue entries.

Data Quality

Multi-Signal Duplicate Detection

Distinguishes shared applicant tracking hostnames (e.g. boards.greenhouse.io) from unique company tenant identifiers to prevent false duplicate rejections.

Attended Execution

Two-Tier Human Review Queue

Isolates community proposals in an append-only staging queue (08_Discovery_Queue) with audit logging (09_Change_Log), requiring human curator review before master inclusion.

CHALLENGING OUR OWN CLAIMS

What we assumed, what we tested, and what actually broke.

Developing an AI tool connected to external platforms exposes the gap between paper designs and platform behavior. Here are four authentic breakdowns from our engineering test harness, and how we engineered defenses around them.

API CONTRACT STAGING

When ChatGPT Actions quietly skipped parameterized endpoints.

How platform-specific OpenAPI parser quirks forced contract adjustments.

What we assumedWe assumed standard OpenAPI 3.1 specifications with nullable types and parameterized REST paths (/v1/sources/{source_id}) would parse identically across modern developer tools and ChatGPT Actions.
What actually happenedOpenAI’s schema parser threw silent validation errors on modern nullable syntax and completely skipped parameterized path operations in the action generator, rendering direct source queries unreachable from the conversational UI.
What we changedWe maintained our strict backend contract while generating a specialized, inlined OpenAPI 3.1 compatibility schema for ChatGPT that inlined path parameters and converted nullable representations without altering server-side typing.
DATA MINIMIZATION

Proving the gateway actively rejects candidate personal data.

Why client privacy must be enforced before requests reach backend sheets.

What we assumedWhen users submit a company discovery proposal, conversational system prompts instructing candidates not to paste personal emails or contact details would be respected.
What actually happenedTest runs showed user inputs and synthetic payloads containing candidate contact information would easily slip past chat prompt instructions, creating persistent GDPR exposure if written to backend spreadsheets.
What we changedWe placed an automated personal-data scanner at the Cloudflare Worker gateway that intercepts requests before database access. In test REQ-9a24cc89-6045-4256-9db3-00269a5741d2, the gateway intercepted prohibited email data, returned code: prohibited_personal_data, and preserved backend row counts (Queue at 37, Change Log at 206) with zero data leakage.
DUPLICATE DETECTION

When duplicate detection confused software platforms with employers.

Handling shared applicant tracking system hosts across distinct companies.

What we assumedWe assumed matching newly submitted company domains against existing registry entries would reliably detect duplicate hiring channels.
What actually happenedWhen multiple Czech companies used the same third-party applicant tracking systems (such as Greenhouse or Lever), domain-level matching flagged distinct employers as duplicates because they shared the same root hostname (boards.greenhouse.io).
What we changedWe upgraded duplicate detection into a multi-signal evaluator that extracts tenant path slugs and ATS organization identifiers, distinguishing between the software provider and the actual employing entity.
INSTRUCTION GOVERNANCE

When operational guardrails exceeded OpenAI instruction limits.

Preserving core invariants under strict platform character ceilings.

What we assumedWe assumed we could put full operational governance, confirmation requirements, error taxonomies, and entity definitions directly inside the Custom GPT instruction prompt.
What actually happenedAs safety controls grew, the prompt exceeded OpenAI’s instruction character ceiling. Default truncation risked publishing a live assistant with stripped-away safety boundaries.
What we changedWe moved all business logic, duplicate checks, and validation rules to the server-side gateway. We compacted the Custom GPT prompt down to a tight baseline that enforces five core invariants—privacy, authority, confirmation, idempotency, and truthfulness—letting the server enforce the heavy lifting.

WORKING DEMONSTRATION

Search curated Czech job sources in ChatGPT.

The Czechia Expat Job Scout is public and free to use. Test how a conversational assistant can query an audited registry without collecting your personal information.

CUSTOM TOOLS OVER VERIFIED DATA

Need safe conversational tools over your company’s data?

If your business has valuable records in spreadsheets, Notion, or internal databases, we build zero-trust edge gateways that let assistants answer questions and draft proposals without risking data leaks or database corruption.