The model under the hood.
Astris HR Model is our proprietary workforce intelligence engine. Today it combines transparent, rule-based scoring, meaning-based similarity matching, a map of how skills relate to one another, and AI-written explanations. As we gather real hiring results, it adds models that learn from those outcomes to sharpen every match — all behind one simple request.
Three layers. One interface.
One interface, two speeds. Ask it for a fast, transparent score the instant a match surfaces, or — when you need to explain a decision — let the same match write a short, plain-language rationale on request. Choosing fast or deep is a request you make, not code we rebuild.
Meaning-based matching
Every candidate and job is turned into a numerical fingerprint of its meaning, so we can compare them by substance rather than exact wording. From a pool of a million people, this narrows each job to the 200 most promising candidates in a fraction of a second. Our map of related skills fills in when the wording doesn't line up.
Seven-factor transparent score
Skills (40%), experience (20%), language (10%), distance (10%), schedule (10%), certifications (5%), transportation (5%). The overall score is a weighted average, with extra credit for closely related skills and a small adjustment for employer quality. As real hiring results come in, learned models join in to refine it.
Plain-language explanation
On request, our AI writes a two-to-four-sentence explanation of why a candidate fits, saved with the match so it loads instantly next time. Privacy safeguards keep work-authorization status, date of birth, and exact address out of the explanation. It is generated only when an employer opens a match to learn more — not for every score.
Every component, weighted explicitly.
Astris HR's scoring is deliberate. No mystery weights. No hidden factors. Every component follows fixed, published rules and reports its own sub-score alongside the overall result, so you can always see what drove a match.
The share of the job's required skills the candidate already has, plus partial credit for closely related skills identified through our map of how skills connect.
Years of experience measured against the job's minimum. No experience scores 20; meeting the requirement scores 100.
The share of the job's required languages the candidate speaks at a conversational level or higher.
Straight-line travel distance against the candidate's stated maximum commute. Remote roles score 100 automatically. A same-state match with no exact location scores 70.
How well the candidate's availability and shift preferences overlap with the job's schedule.
The share of required certifications the candidate has on file. If none are required, this scores 100 automatically.
How the candidate's transportation compares with what the job requires, including access to public transit.
One simple request. Many models behind it.
Astris HR Model is built to combine several models from day one, the way a panel of experts reaches a stronger decision together than any one would alone. The rule-based scorer is live now. As each new model finishes training on real data, it joins the panel and starts improving the overall score the same day — no rebuild required.
Our transparent seven-factor score, tuned by industry. It requires coverage of the must-have skills, credits closely related and similarly-worded skills, gives less weight to work history that is harder to verify from abroad, and lets each job emphasize the skills that matter most to it.
Compares candidates and jobs by the meaning of their text, not just matching keywords, so strong fits surface even when they use different words. It accounts for about a third of the combined score.
Estimates how likely a hire is to stay at 30, 90, 180, and 365 days, anchored to official U.S. Labor Department turnover rates for the job's industry. A model that learns from our own hiring results takes over once we reach 1,000 placements.
For each missing must-have skill, it projects how much a candidate's score would rise if they gained it, then names the top three skills to invest in — with a suggested path and an estimate of how long each takes to reach job-ready.
When reviewers rate and correct how we read a resume or job description, those corrections are fed into the next reading as examples. Accuracy improves continuously, with no retraining needed.
When users rate a translated phrase poorly (below three stars across three or more votes), our AI generates a better wording that replaces the default going forward.
A model that learns which factors best predict a good hire from real results. It turns on automatically once we reach 200 recorded outcomes (or 50 within a single industry).
A more advanced model that captures how factors interact in combination, not just individually. It turns on automatically once we reach 500 outcomes (or 100 within a single industry).
A model trained on real results to predict how long a hire is likely to stay, at the 30-, 90-, 180-, and 365-day marks. It turns on automatically once we reach 1,000 placements.
Automatically groups similar candidates together. This powers the “find more candidates like this one” feature and the similar-candidates panel.
Separately from the candidate, every employer carries a quality score from 0 to 100, drawn from Glassdoor and Indeed ratings, how well their past hires have stayed, and caseworker feedback. Based on that score, a match can gain or lose up to five points — better employers get a small but real edge in the ranking, all else equal, while below-average ones are nudged down.
The model improves from real use.
Document reading ratings
Caseworkers, candidates, and employers rate how accurately we read each document, from 1 to 5. A low rating with a correction prompts an automatic re-read that folds the correction in.
Translation feedback
A simple thumbs up or down on translated text. When a phrase averages poorly across several votes, our AI generates a better wording that replaces the default.
Match feedback
Employers rate match quality on a five-star scale. Every rating becomes a real example the models learn from in the next training cycle.
Hiring outcomes
Whether a hire stays at 30, 90, 180, and 365 days, why anyone leaves, and follow-up survey answers all feed back in to improve how we predict retention.
Document processing in 35+ languages.
Astris HR Model reads and understands resumes and job descriptions in any of these languages. It detects and records each document's original language. Foreign credentials and past occupations are kept in their original language and matched to their English equivalents, never discarded.
Model roadmap.
What's live today, what's ready to turn on, and what's next.
- Meaning-based matching live across all candidates and jobs
- Industry-tuned scoring weights with a must-have-skill requirement
- Retention forecast and wage context anchored to official U.S. Labor Department data
- Skill-gap projections with an estimate of how long each takes to reach job-ready
- Matching that recognizes similarly-worded and related skills
- Less weight on work history that is harder to verify from abroad
- AI-generated translation improvements in production
- Document reading that self-corrects from reviewer feedback
- Public API and connectors to common hiring systems
- Fairness monitor that flags adverse impact (the four-fifths rule)
- U.S. and world heat maps of supply and demand
- Drag-and-drop hiring pipeline
- Outcome-trained scorer — turns on at 200 hiring results (50 within an industry)
- Pattern-learning scorer — turns on at 500 hiring results (100 within an industry)
- Retention prediction model — turns on at 1,000 placements
- Automatic refresh of U.S. Census population data
- Short in-app skills tests candidates can take to confirm their abilities
- Matching for moving existing staff into new roles within a company
- Anonymized, cross-employer labor market insights
- Local unemployment and wage-growth data by metro area
- Fair-market rent data to gauge whether relocation is feasible
- Expanded skills library based on the national O*NET occupation standard
- Dedicated tuning for individual industries (warehouse, hospitality, and healthcare separately)