Technical deep dive
How the matching pipeline works, how the sub-models combine into one score, how we sanity-check the ranking on a synthetic test set, and the public data feeds that set Astris apart from skills-based hiring platforms and from startups that simply hand the job to a large language model.
How we're different
Three commonly-confused product categories. Astris is none of them.
| Capability | Astris HR | Adept ID | Traditional applicant-tracking system | Chatbot-as-matcher |
|---|---|---|---|---|
| Foreign credential parsing (source-language preservation) | Yes | No | No | Partial |
| Skill matching across 35+ languages | Yes | English only | Keyword match | Yes (no review loop) |
| Up-to-date labor turnover data (official government sources) | Yes | No | No | No |
| Retention forecast anchored to the industry, state by state | Yes | No | No | Rough guess only |
| Wage benchmarking (official wage + metro income data) | Yes — every match | No | Manual | No |
| "What if they had skill X" analysis | Yes — top 3 gaps + time to close them | No | No | No |
| Factor-by-factor score breakdown (every match) | 7 factors, each weighted | Not shown | Keyword count | Free text only |
| No high score without covering the required skills | Yes | No | No | No |
| Fairness monitor (flags adverse impact by group) | Yes | External | External | No |
| Re-running a match gives the same score every time | Yes | Yes | Yes | No (answers vary run to run) |
| Self-validation in candidate's native language | Yes | No | No | No |
The matching pipeline
Scoring one candidate against one job takes about 140 milliseconds start to finish. Every step runs in plain code and is recorded, so you can trace exactly how any score was reached.
Each candidate-and-job pair moves through the same stages: pull the skills out of the resume; match them against the job's required skills (exact match first, then close spelling variants, then related skills in the same family); combine seven separate factors into one composite score weighted for the specific industry; cap the score if the candidate is missing skills the job treats as essential; add a meaning-based similarity check that compares the full text of the resume and the job; adjust for how long people typically stay in that industry; and assign a final tier. Every intermediate number is stored, so re-running the same score produces an identical result down to the last digit.
How the sub-models combine
Ten sub-models feed one final score, which is their weighted average. A new sub-model first runs silently alongside the others — scoring, but not counted — and only starts influencing results once it has proven itself.
Internal check on a synthetic test set
We built 70 test profiles by hand, each designed to be a strong fit for one particular job we already knew the right answer for. This is an internal sanity check that the ranking behaves as intended — not a measured real-world result. An independent benchmark on real candidates is still pending. The figures below describe only how the ranker orders these hand-built cases.
These counts come from hand-built test profiles, not real placements. They confirm the ranker orders the intended jobs sensibly; they do not establish real-world accuracy or precision. Validation on a real candidate pool is still pending.
How long people tend to stay, by industry
The same candidate — same stability, same fit — plotted across three industries. Some industries simply have far more turnover than others, which caps how long even an ideal hire is likely to stay.
For the SAME candidate fit, a 1-year retention for food services tops out near 56 %; financial activities tops out near 86 %. Industry physics dominates the long horizon.
We start from the typical odds of staying: if a given share of workers leave each month, we compound that month over month to estimate how many are still on the job after several months. The monthly quit rates come from public government labor data — state-by-state rates from the Census Bureau where available, and national rates from the Bureau of Labor Statistics as a fallback. A candidate's own stability and fit then nudge that baseline up or down, but only within a bounded range (never below about 70 percent of the industry baseline, never above 1.5 times it). Even a perfect candidate in a food-service role has a one-year retention ceiling around 55 percent — the industry itself sets the upper limit.
Five federal data feeds, refreshed monthly
The public data nobody else integrates. All free, all pulled directly from official government sources, and all traceable back to the source.
See Astris HR in your workflow
Run it against your candidate base or a sample JD. We'll set up a tenant in under a day and show you the matches your existing tools are missing.
Request a demo