Work order · Internal · Confidential · CDZ-WO-JL-001
JobLake platform — engineering work order
Build JobLake (working name): every job ad in Australia and New Zealand in one search and one API. Developers and AI agents pay with credits from $5/month; job seekers get a free all-boards search with alerts, and AI CV matching as a $5/month upgrade. This order covers a 2-week launch sprint to Monday 30 November 2026: what ships on launch day, what follows straight after, how to keep jobs fresh, security, spam/abuse prevention, and how we stop customers bulk-copying the data.
Turnstile + email verification + one trial per personLive before public sign-up opens.
§06
08
Daily runs from 18 Nov, with alertsBoard drops > 30% or misses a run by 6 h → alert.
§03
09
Closed jobs never shown as openUnder 2% on the spot-check.
§03
10
Copy test passes at go/no-goA Starter account cannot pull the catalogue.
§09
01Goal & scope
We already collect ~324,000 listings (~303,000 open) across 8 AU/NZ boards: SEEK AU/NZ, Adzuna AU, Careers Victoria, Ethical Jobs, APS Jobs, Prosple NZ and Education Gazette. Turn that into a product people pay for: fresh, deduplicated, safe to sell, and hard to abuse.
Public API (REST), credits, Stripe billing, developer console. MCP server within 30 days
Marketing site (ReviewLake layout), docs, pricing, legal pages (Jobs Index page within 30 days)
Free job-seeker search (alerts within 30 days; $5 Pro with AI CV match in Q1 2027)
Security hardening, spam/abuse controls, data protection against bulk copying, admin & source-health dashboard
Out of scope (this order)
Taking job applications (we always link out to the original board)
Native mobile apps (web app must be mobile-friendly)
Agency lead plan, Clay signals, bulk history licences (Phase 3, separate order)
Countries beyond AU/NZ
Any login-only or paywalled source
02Deliverables — launch sprint & after
Launch date is fixed: Monday 30 November 2026. Anything not needed on launch day moves to the post-launch list. Scope gives way before the date or the security controls do.
W1
17–23 Nov
Week 1: data ready + core platform
Data: daily scheduled runs for all boards from 18 Nov (SEEK every 12 h); normalised schema with employer, location, salary; strip personal data; daily raw snapshots
Dedupe v1 (exact rules: employer + title + location + posted window); open/closed lifecycle; fix the Prosple NZ bug
Accounts: email magic link + Google, orgs, owner/member roles
API keys (hashed) + gateway: auth, rate limits, credit metering
Credit ledger + Stripe: Trial and Starter $5 (Pro/Scale plans can exist in Stripe, but launch marketing leads with $5)
Endpoints: /v1/jobs/search, /v1/jobs/{id}, /v1/jobs/changes (filtered only; full feed = Enterprise)
Data protection core (§07): random + per-org job IDs, page/query limits, hard plan caps, per-org redirect links
Hosting, environments, secrets, Cloudflare in front, CI, backups
W2
24–30 Nov
Week 2: front-end, safety, launch
Marketing site (ReviewLake layout) + /pricing + /sources + /docs (quickstart, reference, credit rules) + ToS / Privacy / Acceptable Use / takedown page
Sweep-pattern and linked-account detection, tracer records (§07); external pen test within 30 days of launch
Q1
2027
Later (separate work order)
Job-seeker Pro $5 (AI CV match, application tracker): needs CV privacy controls built properly, so not rushed into launch
Employer careers pages, more government boards, agency plan, Clay signals, bulk exports, salary insights
03Job refresh — keeping the data latest
The product's promise is "what's open right now". Freshness and correct open/closed status matter more than raw volume.
3.1 Refresh schedule (AEST/NZT, UTC stored)
Source
New-ads sweep
Full status re-check
SEEK AU · SEEK NZ
Every 12 h (06:00, 18:00)
Daily, rolling: 1/3 of open ads per day + all ads closing within 7 days
Adzuna AU
Daily
Daily (pending legal; see §08)
APS · Careers Vic · Ed Gazette · Ethical Jobs · Prosple NZ
Daily
Daily, full (small boards)
Saved jobs & "closing soon"
—
Re-checked every 12 h so alerts never point to a dead ad
Figure 1 — 24-hour refresh cycle (AEST). SEEK is swept twice a day; small boards daily; open/closed status is re-checked continuously.
3.2 How to refresh (incremental, not full re-scrape)
New ads
Crawl listing pages sorted newest-first; stop when a page is entirely already-known IDs
Fetch detail pages only for new IDs (or if the listing summary changed)
Store a content hash per ad; re-parse only when the hash changes → emit job.updated
Polite crawling: per-board concurrency caps, jittered delays, retry with back-off, respect robots.txt where feasible
Open / closed lifecycle
Open: seen in the latest sweep, or detail page live and closes_at in the future
Closed: closes_at has passed, or detail page returns removed/expired, or missing from 2 consecutive status re-checks
Re-opened: a closed ad seen live again → status back to open, event logged
Record first_seen_at, last_seen_at, closed_at; emit job.created / job.updated / job.closed
Figure 2 — Job lifecycle. Every transition emits an event that feeds alerts, webhooks and the changes feed.
3.3 Dedupe across boards
One real job = one job_id with many sources[]. Match on normalised employer + normalised title + location + posted within ±7 days. Exact rules first, then fuzzy score. Scores in the grey zone go to an admin review queue. Target precision ≥ 95% on a weekly 200-record spot-check.
3.4 Freshness targets & monitoring
Metric
Target
Alert when
New ad → in API (median)
< 12 h (SEEK) · < 24 h (others)
> 24 h for any board
Closed ad still shown as open
< 2% of open ads
Spot-check fails 2 days running
Run volume vs previous run
within ±30%
Drop or spike > 30% (likely scraper broken / site changed)
Field coverage: employer · location
≥ 95%
< 90% on any run
Scraper failure detected
< 6 h
No successful run in 6 h past schedule
Each API response carries last_verified_at per job, so customers can see freshness themselves. The public site shows "Updated X hours ago".
04Data model & pipeline
Figure 3 — Data pipeline. Highlighted steps carry launch-blocking controls.
Magic link (single-use, 15-min expiry) + Google OAuth. Session cookies HttpOnly; Secure; SameSite=Lax. 2FA mandatory for admin; optional for org owners.
API keysLaunch
Format jl_live_xxxx…. Store only a SHA-256 hash plus the first 8 chars for display. Shown once at creation. Rotate/revoke in console. Per-key scopes (read, match, alerts) and optional spend cap. Never logged in full.
SecretsLaunch
All secrets (Stripe, DB, LLM, email, proxies) in the platform secret store / env, never in git. Separate keys per environment. Rotation runbook. Pre-commit secret scanning.
Transport & edgeLaunch
TLS everywhere, HSTS. Cloudflare in front: WAF managed rules, bot management, DDoS protection. CORS restricted to our domains for the web app; API is key-based.
Data & DBLaunch
Postgres with least-privilege roles (API read-only on job tables; billing service only on ledger). Encryption at rest. Daily backups, 30-day retention, restore tested monthly. Scraper infrastructure network-isolated from the API/database used by customers.
CVs (personal data)With Pro
PDF/DOCX only, ≤ 5 MB, malware scan, text extracted server-side. Encrypted at rest. User can delete any time. Auto-delete after 90 days inactive. Never used for model training, never shared. Privacy Act (AU/NZ) and GDPR-ready privacy policy.
PaymentsLaunch
Stripe Checkout / Billing only; we never touch card data. Verify Stripe webhook signatures; idempotent handlers; credit grants only from verified webhooks.
Outbound webhooks+30 days
HMAC-SHA256 signature header + timestamp (customers verify; 5-min replay window). Retries with back-off; disable after repeated failures.
App securityLaunch
OWASP Top 10 checklist in code review. Parameterised queries only. Input validation on every endpoint. Output-encode user content. Dependency + container scanning in CI. Automated security scan before launch; external pen test within 30 days of launch.
AI / LLMWith Pro
CV and job text treated as data, never instructions (prompt-injection safe templates). Per-user daily caps. No user data sent to providers that train on it.
Logging & auditLaunch
Structured request logs (key prefix, org, endpoint, credits, latency) with no CV text or secrets. Audit log for admin actions, key changes, credit grants and refunds. Alerts on error spikes.
06Spam & abuse prevention — what and how
Three threats: fake sign-ups farming free credits, people bulk-copying our dataset, and spam or scams reaching job seekers. Items marked Launch must be live and tested before 30 Nov; the rest follow within 30 days.
6.1 Fake sign-ups & trial farming
Cloudflare Turnstile on sign-up, login, contact and alert forms Launch
Email verification required before any trial credits are granted Launch
Block disposable-email domains (maintained blocklist) and +alias repeats Launch
One trial per verified email, Stripe card fingerprint (if card required) and device/IP cluster Launch
Sign-up rate limit per IP / ASN (e.g. 3 per hour); stricter for datacentre and VPN ranges Launch
Trial = 500 credits, 7 days, 2 req/s, no bulk export, no webhooks Launch
Risk score at sign-up (email age/domain, IP reputation, velocity) → high risk = card required +30 days
Stripe Radar + 3-D Secure on paid plans and top-ups Launch
6.2 API abuse & dataset copying
Rate limits per key and per org (token bucket): Trial 2/s · Starter 5/s · Pro 20/s · Scale 50/s, plus daily caps Launch
Credits reserved before the query, committed after. No credits = 402. Too fast = 429 with Retry-AfterLaunch
Page size ≤ 100. Deep offset pagination disabled (cursor only). Full sync only via /changes on paid plans Launch
Non-sequential job IDs (no enumeration by counting) Launch
Anomaly flags: one org pulling > 20% of the catalogue in 24 h; many keys or orgs from one IP/card; sudden 10× usage → auto-throttle + admin review +30 days
Canary records: a few unique, harmless fake-but-plausible fields per org, to trace resold data back to the leaking key +30 days
Acceptable-use policy: no rebuilding a competing job board or reselling raw data; violation = key revoked Launch
Cloudflare bot rules on the public website and search so it can't be scraped in bulk either Launch
6.3 Protecting job seekers
Scam-ad filter on ingest: flag ads that ask for upfront payment, crypto, gift cards, WhatsApp/Telegram-only contact, "no experience, $X,000/week" patterns; flagged = hidden pending review LaunchKeyword rules at launch; v2 within 30 days
We never collect applications. Every job links out to its original board Launch
Alerts: double opt-in, one-click unsubscribe, max 3 free alerts and 1 digest a day +30 days
Email domain with SPF, DKIM, DMARC; bounce and complaint handling; suppress hard bounces LaunchNeeded for sign-in emails from day one
CV upload: type/size limits, malware scan, per-user daily match cap With Pro
07Data protection — stopping bulk copying
A paying customer can always copy what we show them. Our job is to make copying more expensive than paying us, stale within days, and traceable to the account that did it.
Why this works: about 74,700 new ads a week and 41,000 closing within 7 days, so a copy is roughly a quarter out of date within a week. What we sell is the daily stream (fresh, deduplicated, open/closed status, history), not a one-off dataset.
Figure 4 — New ads a one-off copy misses (cumulative, at ≈74,700/week). Within a month a copy is missing about as many new ads as the whole catalogue holds today, and ~41,000 of its ads close every week. Estimate from the 9–12 Nov runs.Figure 5 — Defence in depth. A copier has to get through every layer; the innermost (trace) still catches them if they do.
7.1 API design: never hand over everything
Search returns answers, not dumps: ≤ 20 results per page, ≤ 100 per query, cursor paging only, no "list all" endpointLaunch
Random, non-sequential job IDs: nobody can step through jobs by counting Launch
Per-customer job IDs: each org sees its own opaque IDs, so copies from different accounts can't easily be merged and a leaked file identifies the org Launch
Light by default: search returns summary fields; full job detail costs extra credits Launch
Full daily feed (/changes without filters) and bulk export are Enterprise-only, under a signed licence. Never self-serve Launch
MCP tools return the top 10 per question, built for questions, not extraction When MCP ships
7.2 Make copying cost more than an Enterprise licence
Hard monthly caps on every plan (§08): Starter stops at 5,000 records; top-ups on Starter limited to 2 a month Launch
Even Scale (250,000/month) can't keep pace with ~320,000 new ads a month, so a full mirror needs Enterprise or many accounts (caught by §7.3) Launch
Enterprise full-feed licence: full daily feed + bulk export + history, signed contract, verified company. Price set by Asim (indicative $1,000–3,000/month) Launch
Company checks for Scale and Enterprise: verified business email domain, company details, card on file Launch
Figure 6 — Hard monthly caps per plan vs new AU/NZ ads per month (≈74,700 a week, 9–12 Nov runs). No self-serve plan can keep a full mirror; that needs the Enterprise licence.
7.3 Detect copying
Coverage tracking: the share of all distinct jobs each org has pulled over 7 and 30 days; > 20% on Starter/Pro = auto-throttle + admin alert Launch
Sweep patterns: every category × every location, paging to the end, scheduled bursts. Normal apps and AI users ask scattered questions +30 days
Linked accounts: many orgs sharing a card fingerprint, device, IP range/ASN or email pattern → merged for caps and coverage +30 days
Speed: sudden 10× jumps in usage per org auto-throttled pending review +30 days
7.4 Make leaks traceable
Per-org redirect links: every "apply" URL is go.<domain>/<code>, unique per org, redirecting to the original ad. If our data turns up on another site, its links still run through us: clicks reveal the leaking org, we can switch those links off, and the traffic comes to us. Strongest controlLaunch
Tracer records: a few harmless, org-unique variations (e.g. employer-ID formatting) to prove which account a leaked dataset came from. Never fake jobs that a real job seeker could see +30 days
7.5 Contract & public website
Terms of service: no bulk copying, no reselling or redistributing raw data, no rebuilding a job board or competing dataset. Breach = account closed, links disabled, legal action reserved Launch
Website: Cloudflare bot management + rate limits; results only after a search; no browse-all or sitemap pages listing every jobLaunch
08Credits, billing & legal guardrails
Plan
Price
Hard cap / mo
Limits
Free trial (developer)
$0 · 7 days
500
2 req/s · no export/webhooks
Starter
$5 / mo
5,000
5 req/s · API + MCP
Pro
$29 / mo
50,000
20 req/s · webhooks · filtered changes feed
Scale
$99 / mo
250,000
50 req/s · webhooks · company check · no bulk export
Enterprise full-feed licencesigned contract · verified company
from ~$1,000 / mo
full feed
Full daily feed + bulk export + history · price set by Asim
Job-seeker ProQ1 2027 · not at launch
$5 / mo
fair use
AI CV match (≤100/day) · unlimited alerts · tracker
1 credit ≈ 1 job record or 1 light request; /match = 5 credits. Monthly credits don't roll over; caps are hard (Starter: max 2 top-ups a month); top-ups last 12 months. The credit ledger is append-only and is the single source of truth (grants, usage, refunds, expiries). Every response returns X-Credits-Used / X-Credits-Remaining. Prices are launch assumptions to test.
Legal guardrails (non-negotiable)
Public, logged-out pages only. No accounts, no fake profiles, no paywalled sources.
No personal data from ads. Strip recruiter names, emails and phone numbers on ingest.
Always link back to the original ad; we never take applications.
SEEK and Adzuna: summary fields + link only (no full descriptions) until a lawyer signs off. Confirm whether Adzuna data came via its API; if so, likely no resale, so drop it.
Takedowns honoured within 48 h via /legal/takedown. ToS, Privacy Policy and Acceptable Use live before launch.
Counsel review before the public job-seeker search launches; it competes directly with the boards.
09Launch go/no-go — Fri 27 Nov
Data
All boards ran on schedule every day from 20 Nov (≥ 7 clean days), no missed run > 6 h
Freshness targets in §3.4 met; closed-shown-as-open < 2% on a 100-ad spot-check
Dedupe v1 removes the Adzuna/SEEK double count; precision ≥ 95% on a 200-record sample
Employer + location coverage ≥ 95%; zero personal contact data in the API
Daily raw snapshots present for every run
Platform
New user: sign-up → key → first API call in < 5 minutes using the docs
Credits debit correctly under concurrent load (no double-spend, no negative balance)
Stripe: subscribe, top-up, cancel and failed-payment paths tested in live mode with a real card
Every Launch control in §05–§07 verified; automated security scan clean (high/critical)
API p95 < 500 ms on a load test; uptime monitoring and alerts on
Copy test: a test account on Starter tries to pull the whole catalogue; it is capped, throttled and flagged on the coverage dashboard; its apply links resolve via per-org redirects
If any data or security item fails on Friday, we cut scope (e.g. launch API-only, or drop a board) rather than ship it broken. Asim makes the go/no-go call.
10Timeline & rituals
17–23 Nov
Week 1 · Data + core platform
24–26 Nov
Front-end + controls
27–28 Nov
Freeze · QA · soft launch
30 Nov
Public launch
Date
Milestone
Mon 17 Nov
Kick-off; tech lead confirms stack, hosting and search engine; legal question sent on SEEK/Adzuna
Wed 18 Nov
Daily scheduled runs live for all boards (data clock starts)
Fri 20 Nov
Demo: clean deduped data + first API call with a key and credits
Tue 24 Nov
Stripe live-mode tested; site + docs draft; free search working
Thu 26 Nov
Code freeze; only fixes after this
Fri 27 Nov
QA, security checklist, load test → go/no-go (§09)
Sat 28 Nov
Soft launch to 10–20 beta users; watch logs, credits and abuse flags
Mon 30 Nov
Public launch; on-call rota for launch week
How we run it
Daily 15-min stand-up · demo to Asim on 20, 24 and 27 Nov · blockers raised the same day
Anything not on the launch list goes to the post-launch board. No new scope after 24 Nov
Assumes 3 engineers full-time from 17 Nov, reusing the existing Codeaza Data Platform scrapers. The tech lead flags by Wed 18 Nov if anything on the launch list is at risk.
11Open decisions (Asim)
#
Decision
Needed by
1
Final product name and domain (JobLake?)
Tue 17 Nov (site, emails, Stripe need it)
2
Stack, hosting and search engine (tech lead proposes)
Tue 17 Nov
3
Card required for the developer trial? (less abuse vs fewer sign-ups)
Thu 19 Nov
4
SEEK / Adzuna at launch: summary + link only (default) unless legal clears more
Mon 23 Nov
5
Enterprise full-feed licence price and contract template
Wed 25 Nov
6
Launch channels and announcement (LinkedIn, Product Hunt, AU dev communities, Alexandria application)
Wed 25 Nov
Approval to start
On sign-off the team starts Monday 17 November. Launch is Monday 30 November 2026. Scope changes go through Asim and are logged as amendments to this work order.