The model under the hood.
Astris HR Model is our proprietary workforce intelligence model. Today it composes deterministic scoring, embedding similarity, a knowledge graph, and large language model reasoning. Tomorrow it adds logistic regression, gradient-boosted trees, and a neural network trained on observed placement outcomes — all behind one inference interface.
Three layers. One interface.
Astris HR Model is composed of a recall layer, a scoring layer, and a reasoning layer. The same interface serves a quick deterministic score and a fully reasoned recommendation — the latency-quality tradeoff is a configuration, not a code change.
Embedding similarity
Candidates and jobs are embedded into 1536-dim space. HNSW ANN narrows the field to the top 200 candidates per job in under 50 ms over a million-row table. The skills graph contributes when an exact embedding match doesn't exist.
7-component deterministic + ensemble
Skills (40%), experience (20%), language (10%), distance (10%), schedule (10%), certifications (5%), transportation (5%). Composite is a weighted average with a transferable-skill credit and a company quality adjustment. Ensemble adds trained sub-models as data accumulates.
Plain-language rationale
A large language model produces 2-4 sentence rationales on demand. Cached on the match row. PII gates prevent disclosure of work auth, DOB, or exact address. Generated only when an employer expands a match card — not for every score.
Every component, weighted explicitly.
Astris HR's scoring is deliberate. No mystery weights. No hidden features. Every component runs deterministically and ships its sub-score with the composite.
Fraction of required skills the candidate has, with a 0.6 transferable-credit for skills bridged through the knowledge graph.
Years of experience scaled against the required minimum. Zero experience scores 20; meeting the requirement scores 100.
Fraction of the job's required languages the candidate speaks at conversational+ proficiency.
Haversine miles against the candidate's stated max commute. Remote roles auto-score 100. State match without coordinates scores 70.
Overlap between candidate availability + shift preferences and the job's schedule.
Fraction of required certifications the candidate has documented. Zero required → auto 100.
Heuristic on candidate transport vs. job's transportation requirements + transit accessibility.
One inference interface. Many models behind it.
Astris HR Model is designed as an ensemble from day one. The heuristic scorer ships now. Additional models register themselves into the catalog as their offline training pipelines produce artifacts — and start contributing to the composite score on the same day, with no code changes at the call site.
Deterministic 7-component scorer with skill-graph boost. Reference implementation. Always available.
Cosine similarity over candidate/job embeddings. Wired to the pgvector HNSW index; activates when embedding population completes.
Trained on placement outcomes. Calibrated probabilities. Strong baseline before tree models converge.
XGBoost over the full feature vector. Captures non-linear interactions and handles missing features cleanly.
Multi-layer perceptron for non-linear feature interactions that trees underfit. Smallest model that earns its weight in the ensemble.
Skill / occupation clustering. Surfaces candidate cohorts and powers workforce-planning analytics.
Independent of the candidate, every employer carries a quality score (0-100) derived from Glassdoor and Indeed ratings, our own observed retention, and caseworker feedback. The ensemble adds(qualityScore − 50) ÷ 10to the composite, capped at ±5 points. Higher-quality employers get a small but real advantage in the ranking, all else equal.
The model improves from real use.
Document parse ratings
Caseworkers, candidates, and employers rate every Astris parse 1-5. ≤2 stars + a correction hint triggers automatic re-parse with the hint folded into the prompt.
Translation feedback
Thumbs up/down on translated UI strings. Average below 3 over 3+ votes triggers Claude-generated overrides that win over the static dictionary.
Match feedback
Employers rate match quality on a 5-star scale. Each rating becomes a labeled training row for the next supervised model retraining cycle.
Placement outcomes
30/90/180/365-day retention milestones, termination reasons, and follow-up survey responses are the supervised label for retention model retraining.
Document processing in 35+ languages.
Astris HR Model parses resumes and job descriptions in any of these languages. Source language is detected and stored on every document. Foreign credentials and former occupations are preserved in their source language and mapped to English equivalents without being erased.
Model roadmap.
What we're working on next.
- Embedding pipeline live
- Logistic regression v1 trained on first 500+ placements
- Translation override generator in production
- XGBoost retention model
- Internal mobility matching for employer-owned workforces
- Skills assessment v1
- Neural network (MLP) for non-linear interactions
- Public API + ATS connectors
- Cross-tenant labor market intelligence