ChooseEnterOpenEscBack

Outreach Desk

An autonomous, safety-gated job-outreach control plane — discover leads, verify real emails, and send personalized campaigns under human approval and hard caps.

LivePython · TypeScriptFastAPI · React · Vite · SQLAlchemy · Postgres

Solo — built and operated in production (invite-only, gated behind Cloudflare Access) since June 2026.

The problem

Job outreach at scale is either manual drudgery or reckless spam. The hard part isn’t sending email — it’s a pipeline that finds the right people, proves their addresses are real, keeps a human on every send decision, and never burns sender reputation or crosses terms-of-service lines.

Technical approach

Discovery fans out across Firecrawl, Brave Search, JSearch, Adzuna, and USAJOBS, plus first-class imports (LinkedIn/Handshake/alumni exports, CSV/XLSX) — explicitly no credentialed scraping. Extraction stores source-document evidence per candidate; enrichment builds a durable company-intelligence cache with tiered candidate scoring and visible rationales. Verification runs a self-hosted Reacher instance with a commercial fallback, and only verified work emails are sendable. Sends are batch-approved, daily-capped, suppression-checked, and scheduled with a strict-precedence follow-up decider.

Cost control is first-class: every credential is encrypted at rest, every provider call debits a usage ledger with daily and monthly budgets displayed in provider-native units.

Technical highlights

  • Fail-closed send gates: verification, suppression, caps, and batch approval all sit between a draft and the wire.
  • Async enrichment workers (Redis/ARQ) keep discovery interactive while deep company intelligence builds in the background.
  • 20-provider capability-routed LLM abstraction with encrypted at-rest credentials and budget caps.

Media

The live instance is invite-only; a recorded walkthrough is in preparation. Repository is private (it operates a live sending system).

Why it matters here

For SWE and product audiences: a production system where the engineering interest is in restraint — provable safety gates around an inherently risky automation.

danielju.com

Daniel Ju

I build and operate ML research pipelines, agent infrastructure, and the self-hosted platform behind this site.

RésuméLegal & technical · PDF

Los Angeles, CA · open to relocation

The full site is a 3D menu built for a desktop window, with a keyboard and a pointer. Open danielju.com on a laptop or desktop to walk it.