CVE Core — Recruiting Intelligence Platform

by Christopher Velasco · CVE Sourcing

---

The Problem

Most AI recruiting tools produce inconsistent, generic output — AI slop with a confidence score attached. A "9/10 fit" means nothing if you can't see why, and a sourcing tool that messages candidates without judgment burns your employer brand one InMail at a time.

CVE Core is built around the opposite premise: every score must be auditable, every message must be reviewed by a human, and the rubric — not the model — is the product.

What It Is

A multi-agent recruiting intelligence platform. One orchestration engine runs a registry of specialized subagents — extraction sourcers, domain evaluators, outreach writers — each defined as a persona + model + temperature in configuration, not code. Domain expertise lives in human-readable rubric files a hiring manager can actually read and challenge.

The calibration layer is the moat. We don't compete on profile databases — we compete on judgment: what separates a real semiconductor IP seller from a SaaS quota-carrier with good keywords, scored consistently, with the reasoning shown.

How It Works

LinkedIn / GitHub / Raw Profile Text
        │
        ▼
┌──────────────────────────────────────────────┐
│              CVE Core Pipeline               │
│                                              │
│  Stage 1 · EXTRACT                           │
│  Fast model parses raw text →                │
│  structured candidate record                 │
│                                              │
│  Stage 2 · EVALUATE                          │
│  Deterministic role classifier routes to     │
│  specialist evaluator → scored thesis        │
│  with binary gates + cap chains (0–10)       │
│                                              │
│  Stage 3 · ENGAGE                            │
│  Peer-level outreach drafted from the        │
│  candidate's actual work — human reviews     │
│  and sends. Always.                          │
└──────────────────────────────────────────────┘
        │
        ▼
Scored shortlist · ATS-ready CSV · Reply-tracked funnel

A Model-Agnostic Subagent System

Every pipeline run is a multi-agent execution. Specialized subagents — extraction sourcers, domain evaluators, and outreach writers — are registered in configuration rather than hard-coded: each defined by its persona, model, and routing rules. This keeps the platform fully model-agnostic; swapping an LLM provider is a configuration change, and the structured output contracts never move.

New domains scale horizontally — a new vertical is a matter of persona, rubric, and signal data, not new engineering.

What Makes the Scoring Trustworthy

• Auditable, not a black box. Every score ships with a cited vetting summary, binary gate decisions, and the cap chain that produced it. When the rubric caps a SaaS-only background at 3.5, the output says so and says why.
• Deterministic role routing. A keyword-density classifier maps candidates to role tracks before any LLM sees them — no model gets to decide its own grading standard.
• Rubric-versioned evaluation cache. Edit a rubric and every stale cached verdict is automatically invalidated. Calibration drift is machine-detectable, not vibes.
• Compliance gating built in. The pipeline screens candidate employers against restricted-entity intelligence (e.g. BIS Entity List companies) before any outreach step. Restricted matches are flagged for research visibility and hard-blocked from automated dispatch — contact decisions stay with a human, where they belong.
• Human-in-the-loop, enforced. No outreach module sends anything autonomously. Drafts queue for review; reply data feeds back into rubric calibration.

Domain Coverage

The platform is calibrated for deep-tech and hard-tech talent — domains where generic recruiting tools fail because the signal is technical, the deal motion is unfamiliar, and the talent pool is small. Live deployments span both engineering and commercial / go-to-market hiring across semiconductor and adjacent hardware sectors.

Each domain agent runs standalone — its own rubrics, signal data, and evaluation history — while inheriting the full pipeline. New verticals are deployed as private client engagements; domain specifics are scoped per engagement.

Interfaces

• CLI — full pipeline from terminal: GitHub handle, profile file, or URL in; scorecard out
• MCP Server — conversational operation from Claude Desktop or any MCP client: search, evaluate, classify, and draft outreach in plain language
• CSV Export — shortlists importable into Gem, Juicebox, Greenhouse, Lever, or any ATS

Engagement Models

Two ways to work together:

1. Your team operates it — I build and calibrate the agent for your roles; your team runs it on your infrastructure with your API keys. Candidate data is processed only through the AI providers you approve — never stored or retained by me.
2. Managed sourcing — I operate it for you: scored shortlists delivered weekly, outreach drafted and reply-tracked, rubric continuously recalibrated against real response data. You just interview.

---

Contact

Christopher Velasco · Agent Architect 📧 people@cvesourcing.com 📍 San Francisco, CA

Rubrics, signal data, and evaluation logic in this ecosystem are proprietary. This repository documents the platform architecture; production deployment requires an engagement agreement.