Shortlist India — Phase 1
A daily, de-duplicated shortlist of India-only jobs, scored against your resume,
with every apply link verified before it reaches the report.
Zero npm dependencies. Zero running cost. Node 22+ only (uses built-in node:sqlite and fetch).
Quick start
npm run seed
Probes every candidate board in data/companies.json, keeps the ones that answer, and
records how many India jobs each carries. Run this once, then roughly weekly.
npm run run
The full pipeline. Writes reports/latest.html.
npm run serve
Opens the report at http://localhost:3100 with the Applied toggle persisting to the database.
npm run stats # run history, per-source counts, registry health
npm run reset # clear jobs/matches so everything counts as new again
What Phase 1 delivers
| Capability | Status |
|---|
| Greenhouse / Lever / Ashby / SmartRecruiters adapters | Done |
| Curated + verified company registry | Done — 160 live boards, 201 dead candidates filtered |
| India gate (city aliases, remote-in-India, country codes) | Done |
| Fingerprint de-duplication + near-duplicate repost merge | Done |
| Link verification with trusted-403 handling | Done |
| BM25 + 5-component match scoring | Done |
| Batch 7 report layout, rendered from the database | Done |
| Applied tracking persisted to SQLite | Done |
| CSV export, portal / mode / score / recommendation filters | Done |
Last run: 10,240 postings fetched → 1,071 India → 959 new unique → 60 reported, 0 dead links.
Registry: 160 live boards carrying 1,683 visible India jobs (201 dead candidates filtered out).
Zero-duplicate guarantee, verified: two consecutive runs produced 120 rows with
0 fingerprint overlap — day 2 served the next 60 rather than repeating day 1.
The rule that fixes broken links
Never construct a URL. Only ever store the one the source returned.
Enforced in three places:
src/adapters/* — every adapter reads a provider-supplied field
(absolute_url, hostedUrl, jobUrl, postingUrl). No string assembly anywhere.
A posting with no provider URL is dropped, not guessed at.src/verify.js — every link is fetched and its final redirect target stored.src/report.js — DEAD rows never render.
A 403 from a trusted source means bot-blocked, not dead. Coinbase returns 403 to a
script and 200 in your browser; treating that as death would silently delete good jobs.
Those rows render with an unverified tag.
Scoring
Deterministic, explainable, free. Hover any match % in the report for the full breakdown.
| Component | Weight | What it measures |
|---|
| Must-have skills | 35 | Rarity-weighted coverage of the JD's required skills |
| Semantic fit | 25 | BM25, your resume text against the full JD |
| Experience band | 15 | Your years vs the range the JD asks for |
| Title & domain | 10 | Stemmed overlap with your target titles |
| Location & mode | 10 | Hard gate — a mode you don't accept scores 0 |
| Resume parse health | 5 | Phase 5 — not measured yet |
Phase 1 scores the five measurable components and renormalises to 100 rather than
inventing a value for the sixth.
Why skills are rarity-weighted. "Communication" and "Leadership" appear in nearly
every JD and sit in every PM's skill bank. Unweighted, an upholstery sales role scored a
perfect 35/35. Each required skill is now weighted by how rare it is across the day's
corpus, and a JD offering thin evidence (two generic skills) cannot earn the full
component. That one change moved the off-domain roles out of the top 60.
Bands: ≥75 apply · 60–74 consider · <60 skip.
Configuration
profile.local.json — your search profile. Copy it from profile.example.json:
cp profile.example.json profile.local.json
It is gitignored, so your name and resume text never reach the public repo.
Edit and re-run; nothing is hardcoded.
jobTitles — drives the title component. Add variants you'd genuinely take.skillBank — the truth boundary. Phase 5's resume tailor may only use what's in here.resumeText — free text, fed to BM25. The richer and more honest, the better the scoring.preferredLocations, workModes — the location gate.minScore, dailyLimit — how much reaches the report.sources — which adapters run. Drop one to exclude that portal entirely.excludeKeywords — matched against the title with word boundaries.
(Scanning full JD text dropped 292 real jobs, because plenty merely mention an
internship programme in passing.)
data/companies.json — candidate boards. Add any company; npm run seed verifies it.
Guessed slugs failing here is expected and is the point of verification.
Layout
profile.json your search profile
data/companies.json candidate ATS boards (seed verifies these)
data/shortlist.db SQLite — jobs, matches, runs, applications
reports/ generated HTML, one per run + latest.html
src/
run.js pipeline, stages 01-10
seed.js registry verification
score.js BM25 + component scoring
verify.js link verification
report.js report renderer
serve.js local server for Applied persistence
stats.js run history and registry health
db.js schema + queries
adapters/ Tier B: greenhouse · lever · ashby · smartrecruiters
Tier A: adzuna · careerjet · jooble · jsearch (need keys)
lib/ http · normalize (fingerprint) · india gate · skills · keys
tools/ expand-registry.js
Tier A — reaching Naukri, LinkedIn and Indeed
Tier B (open ATS boards) is keyless but caps out at roughly 1,000 India jobs. Tier A
reaches the Indian portals you actually care about, through aggregators that already
index them — no scraping. Each is free and optional; the pipeline runs fine without
any of them and reports which are idle.
cp keys.example.json keys.json # then paste your keys in
| Source | Get a free key | Reaches |
|---|
| JSearch (Google for Jobs) | rapidapi.com/letscrape-6bRBa3QguO5/api/jsearch | Naukri, LinkedIn, Indeed, Foundit, Shine, TimesJobs, Hirist |
| Adzuna India | developer.adzuna.com | Broad India index, salary estimates |
| Careerjet India | careerjet.com/partners | Broad India index |
| Jooble India | jooble.org/api/about | Broad India index |
JSearch is the one that matters most — it is the legitimate route to the portals in your
original app. job_publisher tells you which portal each row came from, and
job_apply_link is the canonical link Google resolved to, so the no-constructed-URLs
rule still holds.
These four adapters are written to each provider's documented response shape but have
not been executed against a live key — I could not verify them without your credentials.
Expect to shake out a field mapping or two on first run; npm run stats will show any
source warnings.
Known limits in Phase 1
- Tier A is unproven until you add keys. See above. Without them, coverage is open
ATS boards only — good jobs, but not Naukri or LinkedIn.
- The registry rewards expansion. 160 live boards now. Adding companies to
data/companies.json then re-seeding is pure upside; guessed slugs simply get filtered.
- Off-domain roles still creep in around the 75 mark when a JD is generic.
Title scoring flags them (low title points), but true domain filtering is Phase 3.
- Single user, no login.
profile.json is the whole user model until Phase 2.
- Salary is rarely populated — ATS boards seldom publish it. Tier A sources do.
