I am an AI Solutions Architect and System Designer specializing in fault-tolerant cognitive architectures and neuro-symbolic memory systems. My approach is "Architecture First": I design logic, data flows, and failure points before writing code.

Coming from an E-commerce portfolio management background, I build AI solutions with a clear ROI—either by radically reducing infrastructure costs or uncovering hidden profit (e.g., identifying €600k in hidden losses via custom ML pipelines).

A headless search engine built for privacy and efficiency.

• Technology: Rust (Axum, LanceDB, Tantivy) and Python (ONNX).
• Optimization: Used Rust to eliminate memory leaks common in Python-based RAG pipelines.
• Performance: Achieved 91.21% Recall@10 on the HotpotQA benchmark. Runs stable on entry-level hardware (Mac M1, 8GB RAM).
• Link: onecero.one

"Memory-as-a-Service" combining vector search with knowledge graphs.

• Optimization: Rewrote graph physics using LUA scripts inside Redis, reducing latency from 800ms to 40ms.
• Architecture: Implemented strict multi-tenancy and data isolation at the core level.
• Link: myceliummemory.tech

Custom ML model designed to detect Prompt Injection and memory poisoning attacks.

• Model: Fine-tuned a Small Language Model (based on DeBERTa-v3) using custom Red Teaming datasets.
• Metrics: 99.32% Accuracy with a 0.00% False Positive Rate (FPR).
• Deployment: Optimized for CPU inference via ONNX format.
• Link: Hugging Face Repository

An orchestrator for complex, multi-step reasoning using a Planner-Executor pattern.

• Logic: Dynamic agent spawning with deterministic behavior via strict typing (Pydantic) and Structured Outputs.
• Safety: Integrated secure sandbox for isolated Python code execution.

Unsupervised learning engine for high-dimensional market analysis.

• Algorithms: HDBSCAN, K-Means, and UMAP for clustering millions of products and reviews.
• Impact: Audited a €2M+ turnover account and identified €600,000 in hidden inventory losses by bypassing standard ERP reporting errors via custom BigQuery SQL.

• AI & Machine Learning: Python (3.12), Vertex AI, ONNX, PyTorch, DeBERTa/BERT, RAG, Multi-Agent Swarms.
• High-Performance Backend: Rust (Axum), FastAPI, AsyncIO, gRPC, WebSockets.
• Data & Infrastructure: Redis Stack (Graph), LanceDB, Tantivy, GCP (BigQuery, Cloud Run, Firestore), Docker.
• Mathematics: HDBSCAN, UMAP, K-Means, Vector Mathematics, Graph Theory.
• Frontend: React, TypeScript, Vite, Server-Driven UI (SDUI).

• AI Infrastructure: Development of local RAG engines and enterprise-grade agentic memory.
• ML Engineering & Security: Red Teaming, fine-tuning SLMs, and model optimization for CPU inference.
• System Design: End-to-end realization from database architecture to server-driven interfaces.

I am particularly interested in Technical Lead or Founding Engineer roles within AI-focused startups.

• Email: contact@onecero.one
• Telegram: @paulshalyhin
• LinkedIn: Pavel Shalyhin