Deterministic where it should be. Semantic where it helps. Human-approved before anything leaves.
Warpply Core is an async-first application pipeline. It discovers and normalizes roles from structured ATS feeds and curated aggregators, scores fit against your resume profile across four explicit dimensions, and enforces a strict human review gate before anything is tailored or submitted. A match is a scored (user, resume_profile, job_posting) relationship — not a guess.
Volume on one side. Opacity on the other.
The market produces thousands of postings per day with inconsistent titles, ambiguous seniority bands, and vague remote eligibility. Submissions disappear into ATS pipelines optimized for recruiter throughput, not candidate fit. Most aggregators rank by recency and popularity. None of them know what you are actually qualified to do.
Candidate-side scoring matters because nobody else is paid to do it. Warpply ingests from structured sources, normalizes them into a single canonical job_posting shape, and scores every match against your resume profile before it ever reaches a review queue.
Source coverage is additive. Adapters are introduced when a source produces structured, attributable postings — not to inflate the count.
Four dimensions. Hard penalties. Auditable arithmetic.
Every match resolves a (user, resume_profile, job_posting) tuple to a single FIT_SCORE. The base score is composed deterministically from four independent dimensions, then adjusted by penalties that reflect non-negotiable mismatches.
title_role_scoremeasures alignment between the posting's title and the roles your profile already covers. seniority_score compares posted seniority to your stated band. tech_stack_score compares required and bonus technologies to the stack on your profile. location_remote_score evaluates remote eligibility, region constraints, and timezone overlap.
Penalties are applied for eligibility and work-authorization mismatches, salary outside your declared band, weak benefits proxies, and archetype mismatch — for example a long-tenured platform engineer being routed to a short-cycle agency role.
Scores shown are illustrative. Weights and thresholds live in the scorer module and are reproducible per match.
Exact keyword overlap is not enough.
“Postgres” and “PostgreSQL”. “k8s” and “Kubernetes”. “FastAPI” and “Python web services”. A naïve keyword filter rejects half of these as mismatches and routes a qualified candidate to a queue they should never have been in.
Warpply layers an optional MiniLM-style embedding similarity pass over the deterministic tech-stack scorer so related skills are recognized as related, not identical. The semantic pass is additive, never overriding: it can lift tech_stack_score when there is real conceptual overlap, but the deterministic baseline still anchors the score and the trace remains explainable.
LLMs do not score. They adjudicate.
Most matches resolve cleanly with deterministic scoring. A smaller set falls into a gray zone — close to threshold, ambiguous title, atypical seniority phrasing. For those, and only those, Warpply optionally consults a language model to adjudicate the edge.
The LLM is not the base scorer. Its use is bounded by three constraints: it runs only on gray-zone matches, it is daily-capped per user, and it is fallback-safe. If the model is unavailable, slow, or rate-limited, the deterministic FIT_SCORE is the canonical answer and the pipeline keeps moving.
No application is submitted without your approval.
The review gate is architectural, not procedural. There is no setting to disable it. A match must transition through REVIEW_STATUS = APPROVED before any tailoring or submission step runs. Discovery, scoring, and prep all happen async; the boundary between machine work and human consent is the gate.
Automation amplifies judgment. It does not replace it. Warpply will surface candidates, explain the score, and prepare materials — and stop. You decide what goes out.
Operational visibility. Not a black box.
Every stage of the pipeline emits structured telemetry: ingestion counts per source, score distribution per dimension, gray-zone hit rate, review-gate dwell time, and outcomes per match. The data exists so you and we can see how the system is behaving — what got filtered, what got surfaced, what got approved, what got a reply.
This is operational feedback, intended for review and future iteration. Warpply does not claim an automatic model-retraining loop or self-improving inference. Improvements ship the way they always do: when an engineer reads the telemetry and changes the code.
Already scoring jobs with another tool? Warpply is the submission layer for your evaluations — bring a scored job URL directly into the pipeline and inherit the review gate.
Async-first pipeline. Four-dimension fit score. Bounded LLM where it helps. Human review before anything leaves.