PG-AGI — Explainable AI Knowledge Ecosystem
AI Knowledge Ecosystem | Multimodal Product Design | Web & Mobile
CG-AGI SYSTEMS
Role
Lead Product Designer (0 – 1 Platform Architecture, AI Systems Design)
Service
AI Interaction Design, Knowledge System Architecture, High-FidelityUI/UX & Prototyping
Timeline
4 weeks
livelink
Project Overview
When people face high-stakes problems today—like starting a business, navigating taxes, or understanding legal compliance standard AI tools often give vague, unverified answers. At the same time, seasoned professionals like lawyers and consultants have valuable, real-world knowledge, but no practical way to scale it beyond repetitive 1:1 meetings.
I designed PG-AGI to fix this trust gap. The platform converts verified human expertise into interactive, digital entities. It allows users to ask questions naturally via chat or voice and get clear, grounded guidance backed by visible source citations while giving experts an intuitive, zero-code studio to structure and share their knowledge safely.
Problem Statement & Goals
Problem:
When people run into complicated, high-stakes challenges—like setting up a company, handling taxes, or sorting out legal compliance—they turn to search engines and AI for answers. But most tools return generic, unverified text dumps without citing where the advice came from. This lack of transparency makes it risky and intimidating to take action.
At the same time, licensed professionals like lawyers, financial advisors, and consultants possess deep domain knowledge, but they are trapped in repetitive 1:1 meetings with no scalable way to package or distribute their expertise.
Key goals:
- Grounded Trust: Show visible source citations and confidence scores directly in the chat.
- Intent Discovery: Match user problems to verified experts via plain natural language.
- Fluid Multimodal UX: Enable seamless switching between real-time voice and structured chat.
- Zero-Code Studio: Give experts a simple drag-and-drop workspace to publish digital twins without coding.
- Transparent Fallbacks: Surface low-confidence alerts and human audit escalation instead of hallucinating.
User Research & Insights
- To design an interface for high-stakes problem solving, I investigated why users abandon generic AI tools and what prevents domain experts from scaling their knowledge online.
Research methods:
- — Heuristic evaluation of existing Generative AI tools and knowledge platforms to analyze trust deficits, provenance friction, and onboarding workflows.
- —Qualitative interviews with aspiring founders, freelancers, chartered accountants, and legal consultants to map task priorities and mental models.
The Two Core User Groups
- 1. The Problem Seeker (Founders, Freelancers, Professionals): Need actionable, structured steps they can trust for real-world legal and financial tasks, but struggle with generic answers and zero source accountability.
- 2. The Knowledge Provider (Lawyers, CAs, Consultants): Possess deep domain expertise but are bottlenecked by repetitive 1:1 advisory calls and lack zero-code tools to package and monetize their knowledge.
Key Behavioral Insights
- Trust Equals Transparency: Users do not just want fast answers; they need to see who provided the knowledge, the confidence score, and the exact source document.
- Scattered Information Causes Fatigue: Switching between forums, search engines, and chatbots delays critical business decisions and increases cognitive load.
- Non-Technical Experts Need Simplicity: Experts want to scale their advisory capacity without wrestling with complex prompt engineering or node-based logic builders.
User research quote:
- "When I ask a legal or tax question, I don't want a generic essay. I need clear steps and proof of where the answer came from so I don't make an expensive mistake."
Competitive Analysis & Insights
I evaluated direct and indirect players across discovery, conversational grounding, and expert workflows to pinpoint market gaps and design opportunities.
1. Discovery & Selection (Poe vs. Character.ai)
- Strengths: High volume of bots with clear profile descriptors and strong intent matching.
- Weaknesses: Lacks verified credentials; cluttered with repetitive, low-quality bot clones that reduce trust in professional use cases..
- Strategic Takeaway: Quantity does not equal quality in high-stakes problem solving. We need Explainable Ranking that explicitly tells the user why an expert was matched (e.g., "Matched on verified 15-year Corporate Law background").
2. Conversational Grounding (ChatGPT Voice vs. Perplexity AI)
- Strengths: ChatGPT leads in low-latency natural voice interaction, while Perplexity sets the standard for verifiable source citations.
- Weaknesses: ChatGPT's voice mode lacks visual grounding or provenance, while Perplexity's interface is text-heavy and overwhelming.
- Strategic Takeaway: Users need real-time validation without sacrificing interaction speed. We designed a Knowledge Overlay with inline confidence tags and expandable source chips that work across text and voice.
3. Expert Workflow & Ingestion (MindStudio vs. Voiceflow)
- Strengths: Robust data ingestion (PDFs, URLs) and powerful testing playgrounds for real-time conversational tuning.
- Weaknesses: Highly technical node-based environments that intimidate non-technical professionals like doctors and lawyers.
- Strategic Takeaway: To scale, creator onboarding must feel like an intuitive product studio, not an IDE. We prioritized a guided, zero-code ingestion pipeline with tone and formality sliders.
Translating Insights into Design Strategy
- Explainable Discovery: Prioritize intent-based natural language search over generic keyword matching, displaying clear matching rationales on expert cards.
- Proven Grounding & Trust: Integrate inline citation tags and confidence scores (98% Grounded) to eliminate black-box AI hesitation during high-stakes decisions.
- Low-Friction Multimodal Interaction: Enable seamless switching between text and real-time speech-to-speech without losing conversational context..
- Zero-Code Expert Studio: Build a document-driven ingestion workspace with tone and formality sliders so domain experts can deploy digital twins without technical overhead.
- Proactive Gap Alerts: Introduce honest fallback states when datasets are incomplete, offering a direct human-in-the-loop audit option instead of guessing.
User Persona
User flow (suitable for large devices)
Wireframe
UI Design
Screen 1: Discovery Dashboard (Problem-First Entry Point)
Purpose: Shifts the user’s mental model from searching for a static directory of names to solving a context-specific problem.
Natural Language Omni-Search: An expansive search bar that encourages users to describe nuanced, real-world problems (e.g., “I need legal advice for starting a business”) rather than guessing bot keywords.
Intelligence Summary Card: Visualizes query history and highlights relevant domain clusters to maintain conversational continuity across sessions.
Explainable Matching Badges: Surfaces expert entities with explicit match reasons (e.g., “Matched on verified 15-year Corporate Law background”) to establish immediate credibility.
Screen 2: Expert Entities (Transparency & Profile Verification)
Purpose: Provides complete provenance and integrity metrics before a user commits to an advisory session.
Entity Integrity Meter: Quantifies AI representation accuracy (e.g., $9.8/10$) against the expert’s actual published frameworks.
Grounded Knowledge Base: Lists verified, immutable source files (PDFs, logs, case studies) used for reasoning to eliminate black-box hesitation.
Secure Session Initialization: The primary CTA emphasizes end-to-end encrypted sessions to reassure users handling confidential corporate data.
Screen 3: Interaction Workspace (Hybrid Chat & Citations)
Purpose: The core execution canvas where context-aware problem solving happens through structured dialogue.
“Receipts-First” Source Chips: Every response includes clickable citation pills (e.g.,
Q3_Logistics_Report.pdf) answering exactly why and how an answer was formulated.Real-Time Confidence Badges: Persistent indicators (e.g.,
98% Grounded) validate answer reliability in real time.Problem-Solving Stepper: Left sidebar tracks active milestones (Define Objective $\rightarrow$ Upload Context $\rightarrow$ Review Metrics) to prevent conversational drift.
Screen 4: Expert Workspace (Zero-Code Knowledge Structuring)
Purpose: Enables domain professionals (lawyers, CAs) to build and refine digital twins without coding or prompt engineering.
Document-Driven Ingestion: Simple drag-and-drop workspace supporting PDFs, DOCX, and CSV files with clear progress feedback.
Granular Identity Sliders: Non-technical controls for Base Tone (Professional vs. Casual), Formality, and Response Detail to safeguard the expert’s professional voice.
Integrated Playground: A side-by-side sandbox to test and tune AI retrieval before publishing to the directory.
Screen 5: Active Voice Screen (Speech-to-Speech Mode)
Purpose: Delivers a hands-free, low-latency consulting experience for founders and professionals on the go.
Live Dynamic Transcription: Displays spoken words in real time, allowing users to visually verify technical details and reduce voice-only cognitive load.
High-Contrast Minimalist Controls: Large touch targets for mute, end session, and instant switch-back to text mode.
Screen 6: Resilient Edge State (Low-Confidence Alert & Gap Detection)
Purpose: Maintains system honesty and user trust when knowledge base coverage is incomplete.
Transparent System Feedback: Replaces generic 404/error popups with explicit data gap warnings and visual scores (e.g.,
15% Grounding Score).Human Escalation Trigger: Provides a 1-click
Request Auditbutton to route ungrounded queries directly to human domain supervisors.
Screen 7: Expert Performance Analytics Dashboard
Purpose: Gives knowledge providers actionable visibility into their digital reach, grounding reliability, and content usage.
Executive Integrity Metrics: Tracks total consultations ($1,284$), average grounding score ($9.8/10$), and user trust ratings ($99.4\%$).
Most Cited Documents List: Highlights which uploaded reference materials deliver the highest value, helping creators identify high-demand advisory topics.
Key UI Components & Micro-Interactions
Modular Element Library: Reusable components including the omni-search bar, ranked expert cards, and ingestion modules designed for cross-platform scalability.
Provenance & Trust Micro-Chips: Specialized tags for grounding scores (
98% Grounded), verified file receipts, and fallback triggers to keep AI reasoning transparent.Intentional Color Logic: Deep blue anchors platform credibility, high-contrast green validates immutable sources, and amber states highlight knowledge gaps and audit requests.
Strategic Learnings & Product Takeaways
Objective
Trust Is Built on Provenance, Not Generative Speed: In high-stakes domains like legal and financial compliance, users prioritize source verifiability and grounding scores over instant, open-ended conversational text.
Abstracting Technical Complexity for Creators: Domain experts will not scale their knowledge if forced to use code or node-based logic builders. Document-driven ingestion with intuitive identity sliders bridges the creator barrier.
System Honesty Preserves Retention: Acknowledging edge cases through transparent low-confidence alerts and human audit escalations prevents costly errors and builds institutional credibility.
Multimodal Continuity Reduces Friction: Combining real-time speech-to-speech with persistent live transcription allows users to visually verify technical details without breaking conversational flow.
The Future Begins Now
If you would like to work with us or just want to get in touch, we’d love to hear from you!
© 2022 – 2025 | Alrights reserved by David Mathew