Eight steps from article to inbox. Two of them use AI.

EagleEye is an automated outreach pipeline: it finds articles relevant to a client, identifies the author, locates a contact, drafts a personalized pitch, and — if the journalist replies — drafts the follow-up. A human approves every send.

A portfolio rebuild of a real client engagement, rewritten around a fictional outdoor-gear brand, Timberline Outdoor Co. The application runs locally only. This page is the only public-facing artifact.

Schedule a Demo View the code
Demo only — not a live tool. No hosted app, no login, no data collected here.
What it does

The pipeline, and where the model is allowed to touch it

  1. 01

    Search

    Queries news and blog indexes for articles matching a topic brief.

    RULES
  2. 02

    Crawl

    Fetches each result under robots.txt rules and a per-domain rate limit.

    RULES
  3. 03

    Extract author

    Reads JSON-LD, meta tags and byline patterns — structural parsing, no model.

    RULES
  4. 04

    Find contact

    Matches the extracted name against public staff pages and contact records.

    RULES
  5. 05

    Draft the pitch

    Writes a personalized email that cites the specific article and angle.

    AI
  6. 06

    Human review

    Every draft waits in a queue for an explicit approve, edit or discard.

    MANUAL
  7. 07

    Send

    Delivery is hard-locked to a whitelist of test addresses at the code level.

    RULES
  8. 08

    Draft the reply

    If the journalist responds, drafts a contextual follow-up — also reviewed.

    AI

2 / 8

Steps that call a model

0

Emails sent without review

~72%

Author extraction hit rate

Local

Only place it runs

Key design decisions

Four choices that shaped the build

Most of the work was deciding where automation should stop.

01 — Scope

AI scoped to two steps only

The model drafts the outreach email and drafts replies. It never finds authors, never guesses contact details, and never decides who gets contacted — those are deterministic steps you can audit line by line.

02 — Extraction

Non-AI extraction, real failure rate

Authors come from JSON-LD, meta tags and byline patterns. It misses, and the misses are visible: one run returned the outlet Ridgeline Review as the journalist's name, lifted from a byline that read “By Ridgeline Review Staff.” A model would have papered over that with a plausible guess.

03 — Defaults

Automation is off by default

Nothing runs on a schedule. Every run starts from a deliberate manual trigger and stops at hard caps on articles crawled, drafts generated and emails sent.

04 — Safety

Whitelist-enforced sending

Real delivery is locked to a whitelist of test addresses in the send path itself, not behind a UI toggle. A demo cannot accidentally reach a real inbox, however the interface is driven.

Walkthrough

What the local app looks like

Recorded locally · fictionalized data throughout

A ~10-second screen recording: one run from search through approved send.
Dashboard overview
Dashboard overview — runs, usage stats and last activity.
AI draft review
AI draft review — approve, edit or discard before anything sends.
Reply thread
Reply thread — inbound response with the drafted follow-up.
Whitelist management
Whitelist management — the addresses delivery is permitted to reach.
New run form
Starting a run — topic and the article cap for that run.
Stack

Built with

Node.js TypeScript Fastify PostgreSQL Prisma Redis BullMQ Next.js

The app runs locally only and has never been publicly deployed — this page is the single public artifact. Source is on GitHub.

What's real vs. documented

Controls that exist, and gaps that don't pretend to

Built as hard controls

  • +Whitelist enforcement in the send path
  • +robots.txt respected on every fetch
  • +Per-domain rate limiting and per-run caps

Documented as production gaps, not built

  • –CAN-SPAM unsubscribe handling
  • –Data retention and deletion policy
  • –International consent rules (GDPR, PECR)

Want the walkthrough in person?

Twenty minutes, screen-shared, running locally — including the parts that fail.

Schedule a Demo View the code