LeadForge

The problem it solves

Outbound prospecting burns 20+ SDR hours a week on research, email hunting, and follow-up tracking. The standard automation stack — Apollo for data, Hunter for emails, Instantly for sending — runs $300–500 a month before a single reply lands. Worse, volume-first automation actively hurts: it floods the funnel with unqualified leads and pattern-guessed emails, and the bounce rate quietly destroys sender reputation — the one asset outbound can't buy back.

LeadForge's founding thesis is written into its scoring engine: 100 real buyers a day beat 10,000 junk leads. Everything optimizes Revenue > Meetings > Replies > Quality > Quantity, in that order.

What I built

A streaming pipeline that runs continuously: ~40 free-source scrapers ranked by a quality-weighted bandit fill a buffer; each processing cycle drains it through noise filtering, dedup, 0–100 revenue scoring, a hot-lead fast lane, zero-cost contact enrichment, LLM composition, and guarded multi-provider sending. Replies come back through an IMAP poller, get classified into 11 intents, and feed follow-up sequences. A live Next.js Command Center shows the funnel, deliverability, source ROI, and the reply inbox in real time — and a five-squad AI build org maintains the codebase itself under human-merge-only governance.

Key features

• Budget flows to what converts — a quality-weighted bandit scores every source on qualified + contactable + decision-maker yield (decision-makers count 2×). Sources that produce measured-zero conversions get starved; proven converters get floored so they can't be accidentally starved.
• 74% contactability at $0 in data APIs — enrichment waterfall: team-page scraping → pattern-email generation → strict DNS-MX validation → live SMTP handshake, with a port-25 circuit breaker for hostile networks.
• AI outages never stop outreach — one LLM gateway with a 4-tier chain (Claude CLI → local Ollama → metered Haiku API → deterministic templates) that is architected to never raise: a down backend degrades quality, never availability.
• Replies handled like a human would — an 11-intent classifier (auto, unsubscribe, not-the-person, referral, no, objection, meeting, pricing, interested, later, unknown) that strips quoted history and signatures in 5 languages before classifying, with precedence rules so an out-of-office beats a co-occurring "sounds good".
• Self-growing infrastructure — auto-discovers new ATS job boards and caches them, maintains a learned query bank with its own explore/exploit bandit (20% exploration), and runs a nightly self-improvement loop over send outcomes.

System internals

Impact — live database numbers

Reliability & safety engineering

• Sender-reputation guardrails: 100 emails/day hard cap · 1 email per domain · send-window gating · cooldowns between sends and batches · an automatic bounce-halt breaker that stops all sending when the bounce percentage crosses its threshold · a persistent suppression list.
• Human-gated by charter: AI drafts, humans approve sends. The agent build org can open PRs but never merge; arming the build engine requires a double opt-in and a hard daily build cap.
• Never-raise LLM layer: every gateway method catches and degrades — composer falls to templates, scorer and intent classification fall to rules. The system has run whole days on rules alone without stopping.
• Non-destructive persistence: lead saves are merge-upserts after a documented incident where a naive INSERT OR REPLACE wiped enrichment fields on every re-scrape. Nothing downstream is ever silently overwritten.
• Failure isolation everywhere: per-source circuit breakers and quarantine (3 consecutive failures → 30-minute bench), bounded retries with backoff on providers, and a port-25 breaker that switches verification to MX-only on networks that block SMTP.