Six components. One Score.

01 — Pipeline overview

Five stages. Crawl to Score.

The pipeline runs continuously. Every stage is observable, every stage is rate-limited, and every stage emits structured telemetry the re-ranker uses to learn over time.

1. 01

   Crawl

   Direct ATS feeds from the major ATS providers (Greenhouse, Lever, Workday, and others). Continuous crawl from source ATS, refreshed near-real-time. Zero aggregator passes, zero scraped duplicates.

2. 02

   Normalize

   Title canonicalisation, structured location parsing, comp-band extraction, and jurisdiction detection for legal and compliance routing — so every downstream stage sees a consistent shape.

3. 03

   Embed

   High-dimensional semantic embeddings tuned for hiring signals, applied to both the JD and a structured view of the candidate profile. Both live in the same space so distance is meaningful in both directions.

4. 04

   Score

   A semantic similarity baseline, refined by a learned re-ranker that combines skill, level, comp, geo, recency and a number of additional signals into a single fit score per role.

5. 05

   Surface

   A per-candidate quality threshold filters out low-fit roles before display. Roles below the bar stay hidden by design — that's the whole point. You can loosen the threshold; you can never be force-fed noise.

02 — Semantic vs keyword

Meaning beats keywords.

A CV that reads Sr. Product Manager · B2B SaaS almost never lists every adjacent title it should match. A keyword index sees the literal string and stops. A semantic model reads the role shape — function, scope, seniority, market — and surfaces the right neighbours regardless of how the title is phrased.

The practical effect: fewer roles, higher precision. A candidate who historically scrolled hundreds of listings a week to find a handful worth applying to sees those directly, plus the ones they would have missed.

03 — Skill verification

Verify in 12 minutes. Six questions.

Anyone can list "Python" on a CV. Tevos Verify is six questions, twelve minutes, time-boxed, supervised, anti-cheat on. Pass and the badge is permanent, attached to that skill, and the re-ranker treats it as a hard signal — not a claim.

Format

Time-boxed, task-based audit per skill, mixing multiple-choice, output prediction, and short live exercises against a fixed rubric.

Integrity

Supervised session with active anti-cheat measures. Optional candidate-toggled recording; recruiters see whether it was enabled.

Outcome

Pass/fail per skill. A pass attaches a permanent Tevos Verify badge to that skill on the profile, plus a stored score the re-ranker can use as a signal.

Recruiter side

Certified skills surface with a verified badge alongside the candidate name and get a rank boost in the recruiter feed. Recruiters can filter to Certified-only.

Ready to verify? Start Tevos Verify →

04 — Voice-AI screening

First screens at human latency.

Today: candidates record short voice screens in-browser, auto-graded against a per-role rubric. The recruiter sees a transcript and a structured score within minutes. The LLM-driven outbound voice agent (live calls, real-time conversation) is on the roadmap.

05 — Boundaries

Lines we don't cross.

A short, deliberate list. These are the lines we don't cross — listed here so you can hold us to them.

• We don't auto-apply to jobs on your behalf. Every application is one human click.
• We don't sell, rent, or share candidate data with third parties for advertising or training.
• We don't match purely on keywords. Bag-of-words ranking has been off the menu since v1.
• We don't take a fee from candidates — ever. The platform is free to use for the candidate side.
• We don't accept paid placement. No company can buy a higher position in your feed.

06 — Under the hood

For the technically curious.

The full technical specification — GNN architecture, DWO weight optimisation, SEMH ranking, skill graph schema, and API documentation — is available on request for investors, technical press, and enterprise partners.