VineAI Investigator: A Mobile-First $0 Crypto Intelligence Agent Built with Gemma 4

This is a submission for the Gemma 4 Challenge: Build with Gemma 4

What I Built

VineAI Investigator is a mobile-first AI crypto intelligence agent designed to help users rapidly evaluate cryptocurrency tokens using real market data and AI-powered reasoning.

The project was built entirely from an Android phone using Termux,without cloud infrastructure, paid development environments or desktop hardware.

A user can investigate a token with a simple command:

investigate

Example:

investigate pump

The system then:

• Pulls live market data from DexScreener
• Calculates a structured risk score
• Evaluates liquidity conditions
• Detects market anomalies
• Produces an AI-generated forensic report

The goal is not to predict prices.

The goal is to reduce noise and provide a fast intelligence layer for traders, researchers, and curious users.

This project demonstrates how powerful open models can be used to build practical financial intelligence tools with minimal resources.

Demo

Example Workflow

User Input:

investigate vdor

VineAI then:

1. Retrieves live token data from DexScreener
2. Runs a custom risk engine
3. Calculates:
   • Risk Score
   • Risk Level
   • Confidence Level
4. Sends structured intelligence to Gemma
5. Generates a professional investigation report

Repository:

https://github.com/TheCyberVine/VineAI-Investigator

Video Demo:

https://youtube.com/shorts/H4mS46CqTW0?si=FFMqfSDzAHGesGSp

Code

GitHub Repository:

https://github.com/TheCyberVine/VineAI-Investigator

Core Compoments:

• main.py
• router.py
• api.py
• memory.py
• tools/dexscreener_tool.py
• tools/scoring_tool.py
• tools/investigator.py

Architecture:

User Input
↓
Router
↓
DexScreener API
↓
Risk Engine
↓
Gemma 4
↓
Investigation Report

How I Used Gemma 4

Gemma 4 is the reasoning engine behind VineAI Investigator.

Throughout development, I primarily worked with both the Gemma 4 26B MoE model and the Gemma 4 31B Dense model through OpenRouter. Rather than treating them as interchangeable models,I used them for different strengths and trade offs that became apparent during real world testing

The Gemma 4 26B MoE model provided a faster and more responsive experience during rapid iteration and experimentation. This made it particularly useful while developing features, testing prompts, refining investigation workflows, and validating the overall user experience.

For deeper analytical tasks, I preferred the Gemma 4 31B Dense model. In my testing, the dense model consistently produced stronger reasoning, more structured explanations, and higher-quality outputs across technical domains such as programming, cybersecurity concepts, mathematical reasoning, and forensic-style analysis. Since VineAI Investigator is designed to transform raw market data into concise intelligence reports, these reasoning capabilities were especially valuable.

One practical challenge I encountered was API rate limiting on the free endpoints. Because of this, I designed the system to remain functional even when a preferred Gemma model was temporarily unavailable. This experience reinforced an important lesson: building reliable AI systems is not only about model quality, but also about handling real deployment constraints gracefully.

To maximize the value of every model call, VineAI Investigator performs structured analysis before invoking Gemma. Live market data is collected from DexScreener, processed by a custom risk-scoring engine, and then supplied to Gemma as contextual intelligence. This allows the model to focus on reasoning and interpretation rather than basic calculations.

The result is a hybrid system where deterministic logic provides measurable signals and Gemma 4 provides the investigative layer that turns those signals into actionable insights.

As the project evolves, one of my future goals is to support more consistent use of Gemma 4—particularly the 31B Dense model—through local deployment or more reliable inference infrastructure, allowing the platform to fully leverage the reasoning capabilities that inspired the project in the first place.

What I Learned

This project reinforced an important lesson:

AI systems become far more useful when models are combined with tools, data, and structured reasoning.

Gemma was strongest when it acted as an intelligence layer on top of real market information rather than being asked to make unsupported assumptions.

I also learned that meaningful AI projects can be built with very limited resources when open models and open tooling are available while significantly enhancing my python and logic reasoning skills

Something extremely noticeable was how Gemma called out or outrightly flagged errors or weak risk calculation while constructing and testing the risk scoring tool, it autonomously overruled false data or gave major doubts concerning suspicious risk calculation.

Future Improvements

Planned future upgrades include:

• Multi-source token intelligence
• Wallet analysis
• Contract risk analysis
• Autonomous investigation workflows
• Web dashboard
• Long-term memory system
• Multi-agent research mode
• Consistent use of Gemma 4 models

The long-term vision is to evolve VineAI from a token investigator into a broader autonomous intelligence platform.

What makes this project personally meaningful is that it was built entirely from a mobile phone.

As a Nigerian college student and determined future tech innovator without access to high-end hardware currently, I wanted to explore how far open models could take a real-world idea.

VineAI Investigator represents that experiment: combining live market intelligence, custom risk analysis, and Gemma 4 reasoning into a practical tool built from minimal resources.