01 / Project case study
AI intelligence system
An AI-assisted system for discovering, researching, and qualifying U.S. importers from fragmented public and trade data.
System stack
02 / Problem
Freight forwarders often spend hours manually searching for importers, checking websites, reviewing trade signals, finding decision makers, and writing outreach messages. The process is fragmented, repetitive, and difficult to scale.
03 / Insight
A long database of importers does not reduce the hard part of the job. The useful unit is a qualified opportunity: a company with explainable trade evidence, a relevant decision maker, and enough context to support a credible next action.
04 / Product Decision
05 / Workflow
Each stage produces evidence or a decision that the next stage can use. Human review remains explicit rather than being hidden behind automation.
Input request
Find candidate importers
Research company
Analyze trade opportunity
Find decision makers
Generate personalized outreach
Human review
Export result
06 / Architecture
The system separates interface, API, persistence, orchestration, and deployment responsibilities without inventing infrastructure the current product does not use.
Frontend
01Bilingual workflow UI, review states, and result handling.
Backend
02Workflow endpoints, validation, and service coordination.
Data
03Persistent records plus short-lived workflow state.
Agents
04Bounded capabilities connected through explicit evidence.
Provider
05Swappable Fake and external providers keep workflow testing separate from model quality.
Infrastructure
06Local service composition and current deployment work.
07 / My Contribution
My contribution is the product and coordination layer that turns the business problem into a working, reviewable system—not a claim that every line of low-level code was written alone.
AI-native building and product ownership
08 / Product Evidence
Synthetic demo data · Fake Provider · No email sent · Engineering evaluation, not customer production results.
Opportunity workspace
Feature: Workspace and research entry
Internal Beta · repository demo data
The bilingual local workspace exposes Fake mode and keeps the research-to-review workflow explicit.
Qualification result
Feature: Evidence-backed qualification
Engineering validation · Fake Provider
A completed synthetic-company evaluation shows the score, confidence, completeness, recommendation, and supporting signals.
Draft review
Feature: Human review before approval
Internal Beta · no email sent
The generated outreach draft remains gated by a named human reviewer and is never sent automatically.
Website research claims
Feature: Evidence-backed website research
Internal Beta · Fake Research workflow
Public Klein Tools pages are converted into source-linked claims with confidence and Accept, Edit, or Reject controls.
09 / Current Result
Engineering validation
Completed for the v0.2 evaluation set
Ten public companies were engineer-selected and engineer-reviewed to verify safe completion, evidence handling, and the review workflow. This is not customer validation.
Real-user product validation
In progress
The repository contains a validation plan and template, but no real-user results are recorded yet.
What I Learned
Next Step
The next step is to collect honest real-user feedback, improve trade-data evidence, and raise workflow reliability before public deployment or additional autonomy.