Automation

KraftFlow

Evidence-grounded prospecting with deterministic control.

Completed PrototypeAI Workflow / Agent AutomationSolo Product & Automation Engineer
KraftFlow workflow from discovery and evidence capture to human review
Detailed vertical map of the KraftFlow automation system
The friction
1Lead Discovery2Content Extraction3Evidence Gate4Browser Capture5AI Audit6Reference Validation7Human Review
Judgement, made visible.
Problem

Why this needed to exist.

Traditional prospecting tools optimise for message volume rather than relevance. They often generate generic outreach that cannot be justified through real observations. KraftFlow was designed to determine whether a company is genuinely worth contacting, collect traceable evidence and stop unsupported AI claims before they reach a human reviewer.

Context
Internal Tool
Platform
Local Web Application · Background Worker
System

Complexity, given a direction.

  1. 1Lead Discovery
  2. 2Content Extraction
  3. 3Evidence Gate
  4. 4Browser Capture
  5. 5AI Audit
  6. 6Reference Validation
  7. 7Human Review
Decisions

The work behind the interface.

Deterministic control around the LLM

Code manages orchestration, limits, state transitions, validation and safety rules. The LLM is limited to interpreting captured evidence and generating language.

Critical workflow behaviour remains predictable and testable.

Evidence-linked AI findings

Audit findings are rejected unless they reference an extracted evidence record, captured browser source or screenshot.

Unsupported claims are reduced and findings remain auditable.A valid reference does not guarantee that the model interpreted the evidence correctly.

Staged evaluation funnel

Low-cost extraction and filtering run before browser capture and visual AI analysis.

Expensive browser and visual audit operations are limited to stronger candidates.

Evidence, not theatre.

Completed a working local-first prototype covering lead discovery, evidence collection, AI-assisted auditing, scoring, draft validation and human review.

01

End-to-end workflow completed

02

Dedicated worker and controlled concurrency

03

Evidence-linked AI findings

04

No autonomous email-sending capability