Ethical AI Stack · Production Work
In 7 years, we scaled Poker Scientist from a technical experiment to a profitable SaaS with 500 paying users ($29-75/month). As Co-founder and Technical Lead, I solved what the poker industry considered computationally impractical: making Nash equilibrium strategies understandable for human players.
I/O System is a philosophically grounded memory framework that enables Large Language Models to acknowledge uncertainty and limitations without breaking user experience. Unlike most AI systems that pretend to have continuous memory, I/O System implements "Presence over Continuity"—enabling AI systems to be honest about what they don't know through stateless, context-aware interaction patterns. Developed over 4 years with 111 functional requirements across 20 epics, using cognitive-memory for persistent storage while maintaining ethical boundaries.
How do you build AI that remembers without hallucinating—and doesn't cost a fortune? Cognitive Memory solves this with Hybrid-RAG architecture combining semantic search with episodic memory, enhanced by Dual-Judge Evaluation (Cohen's Kappa >0.70 - the gold standard for AI reliability). Serves as the memory layer for both i-o-system and agentic-business, with semantic-memory bridging static PDFs to dynamic memory. Achieves 90% cost reduction (€106 → €5-10/month) through intelligent caching while maintaining quality. Unlike typical approaches, uses scientific validation to prove it actually works.
View ProjectHow can PDFs with curated knowledge be efficiently ingested by LLMs when naive chunking breaks semantic boundaries? Semantic Memory solves this with 3-Level Chunking (Document → Section → Paragraph) that preserves meaning through intelligent boundary detection. Bridges static PDFs to cognitive-memory's dynamic memory layer, enabling i-o-system and agentic-business to access curated knowledge without semantic fragmentation. Most document processing systems use fixed-size approaches that destroy context.
View ProjectHow can AI agents operate in business-critical contexts without hallucinating when existing guardrails are reactive rather than preventive? Agentic Business solves this with 6-Module Hallucination Protection Framework implementing proactive guardrails: Input Validation, Fact-Checking, Consistency Verification, Output Sanitization, Confidence Scoring, and Human-in-the-Loop Escalation. Uses cognitive-memory for reliable persistent storage and i-o-system's ethical framework. Validated with 786 tests ensuring hallucination detection and prevention.
View Project
