Official company and startup career feeds
PUBLIC ENGINEERING CASE STUDY
An evidence-first job-search operating system.
I built a continuously scheduled pipeline that finds fresh Australian data roles, validates the vacancy, ranks opportunity quality, creates truthful application assets and keeps every human decision visible.
Python · pandas · SQLite · Google Sheets API · Streamlit · Caddy · PowerShell · Telegram
END-TO-END WORKFLOW
From a fresh vacancy to a human-approved application.
The pipeline runs continuously on a Windows VPS. Each stage produces auditable evidence for the next, reports partial failures honestly and stops before any external application or message is submitted.
Primary and secondary job-market discovery
All eleven stages, Sydney time
Refresh, organise, report and plan
Official career intelligence
Reads allowlisted Greenhouse, Lever, Ashby, SmartRecruiters and verified portfolio feeds.
Indeed + LinkedIn
Runs exact, location-aware searches for the approved role families and title variants.
SEEK + Jora + Adzuna
Adds independent source coverage with bounded retries, timeouts and source-level metrics.
Clean and validate
Rejects expired, duplicate, senior, ineligible or incomplete vacancies and validates the live JD.
Rank and tailor
Scores the correct base resume, then builds a truthful package only where it can improve the match.
High-match alerts
Sends time-sensitive Telegram alerts for qualified fresh jobs, without applying automatically.
Day 3 / 7 / 14
Surfaces due follow-ups from recorded application dates while preserving manual message approval.
Sort the CRM
Colours and orders the Sheet by expiry, status, freshness, priority and application readiness.
Career Agent + sprint
Creates the daily apply, tailor, follow-up, contact and interview-preparation scorecard.
Market intelligence
Builds employer, semantic-fit, startup, portfolio, research and outcome evidence locally.
Contact intelligence
Ranks up to five verified decision makers and prepares bounded, approval-only referral drafts.
Fresh vacancies lead the queue; the sprint highlights approved roles at ≤24 hours.
Seniority, work rights, location, employment type and mandatory experience are checked before tailoring.
The higher-scoring truthful resume wins; unsupported skills or experience are never invented.
Rahul reviews the resume, submits the application and approves every contact message.
CONTROL PLANEGoogle Sheets CRMPrivate dashboard
APPLICATION ASSETSTailored resume PDFCover letter + evidence
OBSERVABILITYTelegram run reportLogs + source metrics
HUMAN LOOPReview → apply → update status
Outcomes return to follow-ups, analytics and the next daily plan.
ENGINEERING EVIDENCE
Built for failure, not just the happy path.
External job sources are noisy and transient. The design makes degraded sources explicit, persists run state before notification delivery and keeps automation recoverable.
Freshness-first
Jobs are timestamped, tiered and sorted so newly posted opportunities cannot be buried by older inventory.
Truthful ATS tailoring
Resume language is constrained to candidate evidence; missing requirements remain visible instead of being invented.
Operational resilience
Restartable Windows services, bounded timeouts, atomic artifacts and source-level run metrics support unattended operation.
Private by default
The command centre, job records, resumes, contacts and credentials remain behind authentication. This case study contains no private pipeline data.
DESIGN DECISIONS
Automation accelerates judgment; it does not replace it.
No auto-apply
A human reviews every resume and submits every application. This prevents low-quality volume from outrunning evidence.
No automatic cold outreach
Contacts must be professionally relevant and verifiable; messages remain approval-only and bounded.
Observable degradation
A partial source failure is reported as degraded—not silently presented as a successful full-market scan.
AI with deterministic controls
AI-assisted drafting is surrounded by fixed gates, validation, tests and explicit safety rules.
LET'S TALK