PortfolioNavigator - Detailed Analysis | Andrei Zitti

The Challenge

Government officials at BMZ were reviewing dense 200+ page PDF reports, with a large share of their time going into manual document parsing instead of strategic decision-making.

Critical insights were buried deep in bureaucratic text, slowing down vital development aid decisions.

The key insight: a large share of manual review could be reduced with AI document intelligence, if the output stayed reviewable and trustworthy.

12 Weeks of User-Driven Iteration

Week 1

Discovery

User research with BMZ ministry officials. Understanding their real workflows with 200+ page reports.

Week 3

First Prototype

Built a chatbot interface. Technically sound, but completely missed their actual needs.

Week 6pivot

The Pivot

Critical insight: officials needed instant visual synthesis, not conversational AI. Redirected the entire project.

Week 10

MVP Validated

Weekly testing kept us honest. Every feature justified by real user needs, not assumptions.

Week 12

Production Delivery

Shaped by 12 rounds of continuous dialogue with the people who use it daily.

Technical Implementation

Human-Centered Design

• Continuous discovery with ministry officials
• Shifted from conversational AI to visual synthesis after testing
• Co-created wireframes with non-technical staff
• Weekly usability tests and retrospectives

Full-Stack TypeScript

• React + TypeScript frontend
• Node.js + Express API
• PostgreSQL + vector search
• Sub-second RAG query responses

AI Pipeline

• HuggingFace SBERT embeddings (~94% internal retrieval benchmark on German docs)
• Custom fine-tuned summarization models
• Smart chunking + contextual memory for multi-document reasoning

Production Infrastructure

• Dockerized micro-services with CI/CD
• End-to-end monitoring + security compliance
• Deployed in government environment with strict data-privacy protocols

Results & Impact

~90%

estimated review-time reduction

30→3

pages per report

85%

internal pilot adoption

4.6/5

reported user satisfaction

"PortfolioNavigator changed the way we read — turning information overload into instant understanding."

— Internal final-review feedback

Key Learnings

Design thinking is continuous: every iteration aligned the product closer to actual use.
Failures guided the pivot: the wrong prototype was the right lesson.
AI works best when invisible: value came from clarity, not the model itself.
Collaboration drives adoption: by co-creating, users became advocates.

Deep Dive:PortfolioNavigator