Evidence, not assumptions
Method and accuracy
- Pull every engineering and product role, with its full JD, from the board's public API (throttled and cached; refreshed incrementally).
- Regex and heuristics extract countries, visa, legal requirements, years, timezone, onsite days and pay. An LLM reads the eligibility-relevant sentences of every JD with a strict JSON schema.
- Every LLM quote must appear verbatim in the JD or it's dropped. When regex and LLM disagree, the field is marked conflicting and can never make a role “eligible”.
- Verdict: eligible only when each check has evidence. Unclear means the posting doesn't say. It is not a yes.
Accuracy
0 false “eligible” verdicts out of 31 on 50 checked JDs × 4 profiles (India, open to relocating with a visa, 3y; India, not relocating, 3y; US citizen, 5y; Germany, open to relocating with a visa, 6y). It found 86% of truly eligible pairs; the rest show as unclear.
Per field: precision = when a value is extracted, how often it's right. Recall = how often a value stated in the JD is found.
| field | stated in | precision | recall |
|---|---|---|---|
where you can work | 47 | 100% | 94% |
visa sponsorship | 6 | 100% | 100% |
relocation help | 1 | 100% | 100% |
citizenship / clearance / ITAR | 7 | 100% | 71% |
min years | 31 | 100% | 100% |
max years | 4 | 100% | 100% |
onsite days / week | 8 | 100% | 75% |
timezone requirement | 5 | 100% | 100% |
pay in text | 19 | 100% | 100% |