Hello Google AI / Gemini community,
I wanted to share a more complete summary of my recent hands-on experiment with Google AI Studio and Gemini. This started as a simple test, but it quickly turned into a surprisingly advanced application prototype.
The project is called:
COGNILAB OS
AI Sales • Psychology • Persuasion Engineering • Decision Intelligence • Economics • Assistant AI System
The goal was to see how far Gemini could go in building a real, mobile-first web application focused on sales intelligence, psychology, persuasion, decision analysis, and practical AI-assisted workflows.
Overall, I have to say: the result was much stronger than I expected.
Gemini was not just generating small snippets. It helped create a full application structure with branding, navigation, multiple functional modules, a dark/cyan premium interface, export tools, and a product-like layout that already feels close to an MVP.
What we built so far:
1. COGNILAB OS main interface
The application now has a consistent identity and visual style:
- dark futuristic UI,
- cyan/turquoise accent color,
- mobile-first layout,
- module cards,
- hamburger menu,
- section-based navigation,
- global footer with author/publisher information.
Footer branding was added as:
© 2026 COGNILAB. Developed by Mateusz Śliwiński / V3XAM SVN.
AI Sales • Psychology • Persuasion Engineering • Decision Intelligence • Economics • Assistant AI System
2. Sales and decision modules
The application now includes multiple practical modules, including:
- Sales Dashboard
- Bayes Sales module
- Objection Analyzer
- Decision Maker Radar
- Follow-up Generator
- Conversation Simulator
- Call Review
- BATNA / ZOPA Negotiation module
- Coach Mode
- Expert AI / Consultation mode
- Knowledge Base / Cheat Sheet structure
The strongest part so far is that the app is no longer just a static UI. It started becoming a practical sales and decision-support system.
3. Bayesian Sales module
One of the key modules is “Bayes Sprzedażowy” / Bayesian Sales.
The idea is simple: instead of guessing whether a sale is likely to close, the user selects a baseline probability and then updates it based on signals from the client conversation.
Examples of positive signals:
- client asks about price,
- client asks about delivery/start date,
- client asks about contract details,
- client involves a decision-maker,
- client returns with additional questions,
- client compares the offer with competitors.
Examples of negative signals:
- client avoids specifics,
- no clear decision date,
- no decision-maker present,
- strong price objection,
- no response to follow-up,
- no urgent business need.
The module uses an odds / likelihood ratio model:
prior odds → likelihood ratios → posterior probability
Then it returns:
- final probability,
- interpretation,
- recommended next sales action,
- practical sales guidance.
This was one of the most interesting parts of the project because it turned a psychological/sales concept into an interactive decision tool.
4. File export and download fixes
At first, file generation existed visually, but download behavior was unreliable, especially in mobile browsers.
We added and improved:
- Copy to clipboard,
- TXT export,
- HTML export,
- fallback behavior for mobile,
- visible/manual fallback options,
- Blob-based file generation,
- URL.createObjectURL,
- safer delayed URL.revokeObjectURL cleanup.
This was especially important because testing on Android/Brave showed that downloads can behave differently than in a desktop browser.
5. PDF / image export compatibility issue
One of the first important technical issues was related to unsupported CSS color functions.
The app initially generated CSS using modern color functions such as:
- oklab()
- oklch()
- color-mix()
- linear-gradient(in oklab, …)
This caused errors such as:
PDF Generation Error:
Attempting to parse an unsupported color function “oklab”
and later similar issues with image generation.
The fix was to replace unsupported CSS Color Level 4 functions with safer formats:
- HEX
- RGB
- RGBA
- standard CSS gradients
After this, the export workflow became more compatible with PDF/image generation tools.
6. Vite / React / TypeScript fixes
During the process, we also fixed several build/runtime issues.
Examples:
- Duplicate ref issue:
chatEndRef was declared more than once.
Fix:
We replaced duplicate refs with unique names such as:
- consultationChatEndRef
- simulatorChatEndRef
- ReferenceError:
Cannot access ‘calculatePosterior’ before initialization
Fix:
We converted key functions to proper function declarations or moved them before first use.
- Invalid character issue:
ERROR: Unexpected “�”
Fix:
We cleaned the corrupted character from App.tsx.
These were useful real-world debugging cases because they showed that Gemini can build a strong prototype, but generated code still needs proper functional QA and build validation.
7. Navigation and functional QA
Another important step was moving the app toward one central navigation pattern:
activeSection
The goal was to prevent “dead UI” — buttons that look clickable but do nothing.
We focused on:
- real elements,
- working onClick handlers,
- Back to Start buttons,
- working module cards,
- menu navigation,
- mobile tap behavior,
- avoiding overlays that block clicks,
- pointer-events rules for decorative layers.
This was a big lesson: a beautiful AI-generated UI is not enough. Every visible button has to do something real.
8. Mobile and Brave testing
Most of the testing was done on Android, including Brave browser.
That revealed practical issues:
- downloads may behave differently on mobile,
- hidden download clicks are not always reliable,
- PDF/blob handling can be inconsistent,
- UI contrast matters more on a phone screen,
- bottom browser/UI panels can cover controls,
- large decorative fonts can reduce readability.
Because of that, the next stage is UI polish and product hardening:
- better text contrast,
- safer mobile spacing,
- clearer menu items,
- better footer readability,
- stable export fallbacks,
- full click testing.
9. Product direction
At this point, COGNILAB OS feels like more than a demo. It is becoming a prototype for an AI-assisted sales and psychology operating system.
The target concept is:
A mobile-first AI assistant for:
- salespeople,
- sales trainers,
- negotiators,
- people learning persuasion,
- people analyzing conversations,
- people working with decision-making under uncertainty.
The app combines:
- sales logic,
- practical psychology,
- persuasion engineering,
- behavioral economics,
- Bayesian reasoning,
- negotiation tools,
- AI-assisted follow-up generation,
- coaching and analysis modules.
10. What impressed me about Gemini
The most impressive part was not that Gemini generated one perfect file immediately. It did not.
The impressive part was that Gemini was able to work iteratively:
- generate an initial product structure,
- accept corrections,
- fix build errors,
- adapt modules,
- improve UI,
- add export tools,
- implement Bayesian logic,
- respond to functional QA requirements.
In other words, Gemini was useful not only as a code generator, but as a development partner during iterative prototyping.
11. What still needs improvement
The main areas where Gemini-generated applications could improve:
- stronger validation before returning code,
- avoiding unsupported CSS functions in export-ready projects,
- better mobile/browser compatibility assumptions,
- fewer dead UI elements,
- stronger checking for duplicate refs/variables,
- safer file download patterns,
- clearer separation between visual design and actual functionality,
- better build-aware output for React/Vite/TypeScript.
The biggest lesson:
Gemini can create an impressive prototype very quickly, but real product readiness still requires:
- functional QA,
- mobile testing,
- export testing,
- security review,
- API key handling,
- privacy policy planning,
- Play Store readiness checks.
12. Current status
The current version of COGNILAB OS already includes:
- product branding,
- advanced UI,
- multiple sales/psychology modules,
- Bayesian sales analysis,
- export/copy options,
- mobile-first design,
- author/publisher footer,
- a roadmap toward Play Store readiness.
The next stage will be:
- UI polish,
- contrast improvements,
- mobile hardening,
- final QA,
- export reliability,
- security review,
- privacy policy,
- app icon and store listing preparation.
Summary:
This was one of the most interesting AI-assisted development experiments I have done so far. Gemini in Google AI Studio showed strong potential for building real application prototypes, especially when guided with precise technical instructions and iterative feedback.
COGNILAB OS is not yet a finished production app, but it is already a serious MVP-level prototype with real product potential.
I am sharing this as feedback and as a real-world example of how Gemini can be used not only for simple code snippets, but for building, debugging, refining, and shaping an actual AI-assisted application concept.