We Don't Just
Prompt & Pray.
PrediktFi uses a Neuro-Symbolic Architecture. We combine the reasoning power of GPT-5.2 with rigid, deterministic code that enforces investor discipline.
The Engine
Reasoning Layer (Probabilistic)
Model: GPT-5.2 (Reasoning)
We utilize an early-access reasoning model (GPT-5.2) capable of "System 2" thinking. Instead of next-token prediction, it simulates an investment committee debate before outputting a score.
> Analyzing market saturation for 'AI wrapper'...
> Querying product hunt trends & git activity...
> DETECTED: Core logic is just an OpenAI API call...
> VERDICT: Discounting innovation score by 40% (Wrapper Tax).
Why not just ChatGPT?
Standard LLMs are "people pleasers". They hallucinate success. Our system prompt (The Validator) is 2,000+ words of contrarian financial instruction designed to reject 95% of ideas.
Context Injection
We don't just send your text. We inject live Solana price data, SaaS metrics, and Trend Narratives into the context window.
The Hard Rails
Deterministic Layer (Rule-Based)
After the AI scores a project, we run the result through calibration.ts. These are hard-coded rules that override the AI to prevent hype-based scoring.
The "Solo Founder" Cap
If you are a solo founder building a Complex Platform, algorithm caps your Execution Score at 60/100.
maxScore = 60;
flags.push("Cap: Solo Founder Risk");
}
The Buzzword Tax
If your SaaS or token pitch relies on generic keywords like "AI-powered" or "Quantum" without technical depth, we deduct -20 points instantly.
score -= 20;
flags.push("Penalty: Generic Narrative");
}
Verification Layer
Fact-Checking (RPC & APIS)
Mint Authority
We query the Solana blockchain directly. If mintAuthority is not null for a traded token, we flag it as High Rug Risk.
Freeze Authority
Can the dev freeze your wallet? We check the account extensions on-chain. Ideal state: Revoked.
LP Lock Status
NOTE: Currently relying on self-reported addresses cross-referenced with Raydium/Pump.fun API patterns.