Verdikt AI — Multi-LLM design reviews that think like senior designers
How I designed and built an AI-powered product design review tool that routes feedback through three large language models — giving designers and startups instant, structured critique on their UI work without waiting for a senior reviewer.
Gemini Pro · Claude Sonnet · LangChain & LangGraph
My Role
Product Designer & Full-Stack AI Developer
Background
Designers needed honest, structured feedback — without relying on availability or seniority
Design feedback is one of the most valuable and hardest things to get consistently. Junior designers often lack access to experienced reviewers who can pinpoint issues in layout, hierarchy, accessibility, and interaction patterns. Solo designers working independently have no one to challenge their decisions. Even within teams, disagreements about design direction can stall progress — and seeking external review is slow and expensive.
Verdikt AI was built to solve that problem. It's an AI-powered product design review tool with a threaded chat interface that simulates the role of a senior product designer. Users upload screenshots of their web or mobile app designs, describe the context and flow, and receive structured feedback — covering pros, cons, accessibility concerns, and an overall evaluation summary.
What makes Verdikt AI different is its multi-model approach. Instead of relying on a single LLM, the tool routes each review through three separate language models — GPT 4.0, Claude 3.5 Sonnet, and Gemini 2.5 Flash — using a LangChain feedback graph. This produces a three-panel review that gives designers multiple expert perspectives on every design, reducing bias and surfacing issues that a single model might miss.
How might we give designers and startup teams instant, structured design feedback that mirrors what a senior product designer would provide — without the wait, the cost, or the scheduling overhead?
App Features
A focused toolset for design critique at speed
Verdikt AI is intentionally scoped around one job: giving designers fast, useful feedback. Every feature serves that goal directly — no bloat, no distractions.
Chat Interface
A clean, familiar chat experience modelled on how designers already interact with AI tools like ChatGPT and Claude. Threaded messages keep feedback organised and easy to revisit across sessions.
Image Upload
Users can attach up to three design screenshots (5 MB max each). The images are processed directly by the LLMs, enabling visual understanding of layout, typography, spacing, and hierarchy.
Contextual Input
A text input lets designers describe the flow and context behind their screens before submitting for review. This ensures the AI evaluates the design within the right product context, not in isolation.
Three-Panel Review
Each review returns a structured summary from three LLMs — displaying Pros, Cons, Accessibility notes, and an overall evaluation side by side. Three models, three perspectives, one clear picture.
UX Approach
How Verdikt AI was shaped
01
Familiar Patterns, New Purpose
The chat interface was chosen deliberately — it's a UX pattern designers already know from tools like ChatGPT and Claude. By reusing that mental model, onboarding is near-zero. Users immediately understand how to submit a query and interpret the response.
02
Multi-Model Architecture
Getting three LLMs to produce consistent, structured output was the biggest technical challenge. Finding the right models, configuring the LangChain feedback graph, and handling deployment edge cases required significant iteration — especially around Claude's API behaviour.
03
Error Handling & Polish
Deployment surfaced several issues — error states between the server and frontend needed clear visual indicators, and PDF report generation had formatting bugs that were identified and resolved through hands-on debugging.
Product Design Patterns
Design patterns that support the AI workflow
The interface is built around three core patterns that keep the experience clean while handling the complexity of multi-model AI responses under the hood.
Sidebar
Thread management
A persistent sidebar stores the user's review threads and past feedback. Designers can return to any previous review, track iterations, and delete threads they no longer need — keeping the workspace tidy.
Loading
Thinking states
When the three LLMs are processing a review, the UI reflects a pulsing 'thinking' animation. This gives designers a clear signal that work is happening — essential when responses from three models take a few moments to complete.
States
Success & error feedback
Every server-side event has a corresponding frontend indicator. Successful reviews land cleanly. Errors — whether from API limits, image processing, or network issues — surface clear, actionable messages rather than silent failures.
Reports
PDF download
After a review is complete, users can download a PDF report summarising the feedback from all three models. This makes it easy to share results with teammates or stakeholders outside the app.
Tech Stack
An AI-native stack built for multi-model orchestration
Verdikt AI's stack was chosen to handle the unique challenge of routing design reviews through multiple AI models while keeping the frontend responsive and the data layer reliable.
1
Next.js — Application Framework
Next.js provides the full-stack foundation — server-side rendering for the marketing pages, API routes for the LLM orchestration layer, and fast client-side navigation for the chat interface.
2
Shadcn UI & Tailwind CSS — Design System
Shadcn components and Tailwind CSS ensured the interface stayed consistent and polished while iterating rapidly on the chat, sidebar, and review panel layouts.
3
LangChain & LangGraph — AI Orchestration
LangChain handles the connection to all three LLMs. LangGraph manages the feedback graph — a structured pipeline that routes each design review through GPT 4.0, Claude 3.5 Sonnet, and Gemini 2.5 Flash, then aggregates their outputs into a unified review.
4
Three LLMs — GPT 4.0, Claude 3.5 Sonnet, Gemini 2.5 Flash
Each model brings a different evaluation perspective. Running reviews through all three reduces single-model bias and gives designers a richer, more balanced critique of their work.
5
Supabase — Database & Auth
Supabase stores all review threads, user data, and feedback history. It also handles authentication, providing a seamless login experience without additional infrastructure.
6
GitHub & Vercel — Deployment
Version-controlled on GitHub and deployed via Vercel's continuous deployment pipeline. Every push to main is live in seconds, enabling fast iteration cycles.
User Journey
From landing page to design feedback in under a minute
The user journey is designed to minimise friction and get designers to their first review as quickly as possible — including a try-before-you-sign-up option that removes the biggest barrier to adoption.
1
Try Verdikt AI
New visitors can try the tool up to three times without creating an account. This lets designers experience the value of multi-LLM feedback before committing — removing the sign-up barrier entirely.
2
Start a New Chat
The 'New Chat' button creates a fresh review thread. Each thread is a self-contained conversation — designers can run multiple reviews across different projects without losing context.
3
Upload & Describe
Users upload their design screenshots and describe the flow and context in the text input. This two-step submission gives the LLMs enough information to evaluate the design within its intended product context.
4
Review & Download
The three-panel review appears with structured feedback from each LLM. Once complete, a PDF report can be downloaded for sharing with teammates or archiving for future reference.
Design Impact
What Verdikt AI delivered in practice
Verdikt AI was put to the test in a real-world scenario — used on active design work to evaluate and improve production UI. The feedback proved immediately actionable, surfacing issues that would have otherwise reached development.
3
LLM perspectives on every design review — reducing single-model bias
Real
Used on live project work to catch typography, contrast, and accessibility issues before handoff
Fast
Design feedback delivered in seconds — replacing hours of waiting for senior review availability
Beyond speed, Verdikt AI proved valuable as a productivity tool for improving design quality. It consistently caught issues with typography, colour contrast, and accessibility compliance — areas that are easy to overlook under deadline pressure but critical to user experience.
Takeaways
01
Gather more designer feedback on desired features. The core review loop works — the next step is understanding what additional capabilities designers want most, whether that's component-level feedback, design system compliance checks, or competitive benchmarking.
02
Redesign the dashboard experience. The current dashboard is functional but minimal. A more polished UI — with better thread organisation, search, and visual hierarchy — would make Verdikt AI feel like a premium daily-use tool.
03
Launch with tiered memberships. The try-before-you-sign-up model validates interest. The next phase introduces paid tiers with expanded usage limits, team collaboration features, and priority access to newer LLM models.