Bitized Qualifier

An AI-powered lead qualification and outreach preparation system built for B2B service businesses targeting midsize companies in the Netherlands.

Bitized Qualifier takes a raw list of LinkedIn Sales Navigator prospects, researches each company through web intelligence, qualifies them against a detailed Ideal Customer Profile, and drafts personalized Dutch-language outreach messages. All with cited sources, so every claim is verifiable before a message is sent.

The problem it solves

Cold outreach at scale has a quality problem. Most automation tools optimize for volume: blast 500 generic messages and hope for 5 replies. That approach damages your brand, burns through prospects, and produces conversations that go nowhere.

The alternative, fully manual research and personalization, produces great messages but doesn't scale. Spending 15-20 minutes per prospect means a solo founder can realistically reach 30-40 people per week.

Bitized Qualifier sits in the middle. It automates the research and qualification work (which is 80% of the effort) while keeping the human in the loop for review, editing, and sending. The result: outreach that reads like it was written by someone who actually looked at your company, at the volume of someone who has a team.

How it works

1. Import. Upload a CSV export from LinkedIn Sales Navigator. The system deduplicates against existing leads automatically.

2. Research. For each prospect, the system runs structured queries through Perplexity's search API across five dimensions: company overview, operational signals, recent news, the person's background, and relevant job postings. Every fact comes back with a source URL.

3. Qualify. Claude evaluates each researched prospect against a detailed ICP definition. The qualification looks for specific operational signals: companies running business-critical processes on a combination of standard tools plus Excel, because the standard tool can't handle their industry-specific rules. Each qualification verdict (Send, Maybe, Skip) comes with cited evidence and reasoning.

4. Deep investigation. Prospects that come back as "Maybe" automatically get a second research pass with deeper web analysis. If the deeper research surfaces operational signals, they get re-qualified. If not, they stay as Maybe for manual review.

5. Draft. For every qualified prospect, the system generates two personalized Dutch-language messages: a LinkedIn connection request (under 200 characters, observation-only, no pitch) and a follow-up message (referencing specific evidence found during research, with a relevant case study). Every message is grounded in verified evidence only. Unverified claims are explicitly excluded from drafts.

6. Review. A web-based dashboard shows qualified leads with their evidence, source URLs, and draft messages. Click a source URL to verify. Edit a draft inline with a live character counter. Approve to copy to clipboard. Then paste into LinkedIn and send.

What makes it different

Cited evidence, not AI confidence. Every factual claim about a prospect's company links to a specific source URL with a quoted passage. If Claude can't find a source, the claim goes into an explicit "unverified" bucket that's excluded from outreach messages.

Operational qualification, not demographic filtering. The system doesn't just filter by company size and industry. It looks for specific operational patterns: ERP implementations with Excel workarounds, job postings that describe manual data work, growth signals that suggest processes are breaking.

Human-in-the-loop by design. The system prepares everything. A human reviews, edits, and sends. No automated LinkedIn messaging or no browser automation.

Cost-efficient at real volumes. Processing 300 prospects through the full pipeline (research, qualification, drafting) costs approximately $13. That's about $0.04 per prospect.

Tech stack

• Next.js with App Router and TypeScript
• Supabase (Postgres) for data storage and real-time progress tracking
• Anthropic Claude API with Batch API for qualification and drafting (50% cost reduction)
• Perplexity Sonar API for web research with cited sources
• Inngest for durable background job processing
• shadcn/ui + Tailwind CSS for the review interface
• Zod for structured output validation on all AI responses

Architecture highlights

The system separates research from judgment. Perplexity handles factual research (what does this company do, who works there, what tools do they use). Claude handles qualification judgment (does this match the ICP, what's the right outreach angle). This separation means each model does what it's best at, and citations flow cleanly from research into qualification into drafts.

Background processing runs through Inngest with a fan-out/batch hybrid pattern. Research fans out per prospect (parallel, independently retryable). Qualification and drafting batch into single Anthropic Batch API requests (50% cost discount). Long-running batch operations use a durable sleep-poll pattern so processing survives restarts.

Evidence integrity is enforced at every stage. Research results store raw API responses alongside parsed citations. Qualification evidence links back to research source URLs. Draft messages are post-validated: every source URL referenced in a draft is checked against the qualification evidence table. Hallucinated references are flagged before the user sees them.

Status

Currently in active development. The v0 (CLI-based qualification engine) has been used to process 100+ real prospects. v1 (web UI, Perplexity research, AI drafting) is in progress.