An automated pipeline that scrapes, filters and scores PM job listings — delivering tiered alerts via email and WhatsApp so I spend minutes, not hours, on job search every day.
100+
Daily listings processed
5
Filter layers
<2 min
Alert delivery time
7→
Iterations and counting
Problem & context
45 minutes a day on a task that shouldn't exist
As an active PM job seeker in Melbourne, I was spending 45+ minutes daily scanning Seek and Indeed manually. The signal-to-noise ratio was low: most listings were physical product managers, pure sales roles, or jobs requiring clinical registration. The process was repetitive, inconsistent, and unscalable.
The goal wasn't to automate the decision to apply — that still requires human judgment. The goal was to eliminate the work that precedes the decision: finding, filtering, and prioritising.
Solution · System design
A pipeline, not a product
The right mental model was a decision-support pipeline, not a job platform. No UI, no database, no user accounts. The fastest path to value was: fetch → filter → score → alert.
Seek+Indeed→Apify scraper→Dedup layer
↓
5-layer hard filter→AI scorerJD extraction → background match → score/11
↓
Google Sheets+Gmail (3-tier)+WhatsApp ≥8
Product thinking · Key decisions
Every rule is a tradeoff
The decisions that shaped the product weren't technical — they were product decisions about scope, cost, precision, and explainability.
01
Apify over a custom crawler
External dependency accepted in exchange for speed and reliability. Building a crawler would take weeks; Apify took hours. The cost is $14/month — acceptable for an MVP.
Speed over control
02
Keyword-first, AI-confirm
Hard filter uses keyword triggers to flag suspicious roles, then OpenAI confirms intent. This eliminates 60%+ of listings cheaply before expensive AI calls.
Cost efficiency over simplicity
03
JD extraction over keyword matching
Scoring extracts 5–7 core JD requirements first, then matches against my real background. Generic skill matching always returned 4/5 — useless. Extraction-based scoring creates actual variance.
Accuracy over simplicity
04
WhatsApp for urgency, email for digest
Score ≥8 triggers WhatsApp immediately. All pass jobs go in daily email, sorted by tier. Two channels, two jobs: real-time alerting vs. structured review.
Engagement over simplicity
Build & iteration
Seven versions in one week
Each iteration fixed a real problem observed in production, not a hypothetical one.
v1
Indeed only · keyword filter
Prove the pipeline works end to end.
v2
Added Seek · fixed actor ID
Seek actor ID had l/I ambiguity. Broader coverage.
Prove the pipeline works end to end. Established the core fetch → filter → notify loop.
Seek actor ID had an l/I character ambiguity causing 404 errors. Fixed and added second data source for broader coverage.
Videographer and WHS specialist roles were passing the filter because companies mentioned "wellbeing." Added OpenAI semantic classification to remove non-relevant roles before the hard filter.
Part-time roles were slipping through. Added FTE conversion (0.53 FTE threshold) and hour-per-week parsing to correctly block over-20h/week roles.
v5 skill match always returned 4/5 — generic PM keywords matched every JD. Redesigned in v6: extract 5–7 core JD requirements first, then match against real candidate background. Scores now reflect actual fit.
Added WhatsApp alerts for score ≥8, 3-tier email layout, weekday-only scheduling, 14-day deduplication window, and crash notification emails.
Reflection · What I learned
Three things that surprised me
1
Hard filter design is product design.
Every rule is a precision/recall tradeoff. Too strict and you miss good roles; too loose and the signal drowns in noise. Writing each filter rule forced me to be explicit about my own constraints and values as a candidate.
2
AI scoring needs a reference point, not a keyword list.
Generic skill keywords match every PM job description — useless. Useful scoring requires a fixed candidate profile as an anchor: real experience, real gaps. The model only became meaningful when it had something specific to compare against.
3
Operational reliability is a product decision.
Scheduling, error notification emails, daily-run-once state, deduplication windows — these feel like engineering details but they're product decisions. A pipeline that silently fails is worse than no pipeline.