Frauenfeld, Switzerland · 47° 33' N 8° 54' E · systems online
Ioannis
Banousis
AI Agent Engineer at Aumera AI working on Agent reasoning, evaluation and optimization.
Building AI that ships.
From advanced mathematics to production AI solutions that drive real business impact.
background
With an Integrated MSc in Applied Mathematics & Statistics from NTUA, I've built my career at the intersection of advanced mathematics and cutting-edge AI technology. My thesis on Econometric Methods of Cryptocurrency Volatility Estimation (R Programming) achieved a 9.5/10 grade.
Currently, working at AumeraAI in Solothurn, Switzerland.
key_achievements
- Deployed 7+ production AI applications with measurable business impact
- Established AI governance and best practices across the organization
- Architected scalable solutions integrating OpenAI, Anthropic, Google, and open-source LLMs
- Previous experience in fraud detection and risk analysis at Kaizen Gaming
Featured Projects
Production AI applications and personal research systems. Click any card to inspect its architecture.
AI Data Analysis Platform
Natural Language Query (NLQ) business intelligence tool that enables non-technical users to query databases and generate visualizations using plain English.
key_features
- Natural language to SQL conversion
- Automated data visualization
- Real-time business insights
- Multi-database support
Predictive Analytics Engine
Advanced forecasting solution designed for F&B businesses, predicting daily/hourly orders, product demand, and raw material consumption.
capabilities
- Time-series forecasting (85%+ accuracy)
- Demand prediction models
- Inventory optimization
- Resource planning automation
Testing Platform
Unified testing platform that orchestrates 6 specialized testing modules, providing comprehensive quality assurance and automated security analysis.
testing_modules
- Lighthouse Performance Auditor
- OWASP Security Scanner
- AI-Powered Automated Tester
- SEO Performance Analyzer
- GitHub Commit Scanner
- W3C/WCAG HTML Validator
Stock Analysis System
Multi-agent AI investment research platform powered by LangGraph and Claude 4.5 Sonnet. Five specialized AI agents analyze stocks in parallel, self-correct through reflection loops and reach consensus through structured debate.
key_features
- 5 Specialized AI Agents (Fundamental, Technical, Risk, ESG, Sentiment)
- Agent Self-Correction Loops (Draft → Reflect → Refine)
- Multi-Round Debate System with Convergence Tracking
- Real-time WebSocket Streaming
- Production-ready FastAPI Backend
PenBot
AI Chatbot Penetration Testing Framework powered by LangGraph, Claude Sonnet 4.5, and Model Context Protocol. Multi-agent security testing system with evolutionary learning, deep reasoning, automatic tool discovery, finding persistence verification, and comprehensive OWASP LLM Top 10 2025 coverage.
key_features
- 13 Specialized Agents (incl. Evolutionary, Token Soup, RAG Poisoning, Tool Exploit, Indirect Injection, Exfiltration, Action Safety)
- Deep Agent Pipeline: Subagent Refinement + Think-MCP Reasoning
- 1,378+ Attack Pattern Templates with Real-time Mutation across 26 libraries
- 22 Vulnerability Detectors: Two-layer detection (pattern + LLM)
- Multimodal Support (Vision/Image Attacks) & Tavily OSINT
- OWASP LLM Top 10 2025 Coverage (9/10 categories) & Attack Lineage Tracking
- Automatic Tool & API Discovery with Persistence Verification
Neural Colosseum
Connect Four LLM Benchmark — a controlled testing framework that pits frontier language models against a minimax-optimal solver to measure rule compliance, strategic depth, and consistency. Supports 14 model variants across OpenAI and Anthropic with real-time visualization.
key_features
- 14 Model Variants: GPT-5.4, GPT-5-mini, GPT-5.2, Claude Opus/Sonnet/Haiku with thinking on/off
- Minimax Baseline: Alpha-beta solver (depth 1–8) with move grading (optimal/good/decent/blunder)
- 4 Benchmark Phases: Rule Compliance, vs Minimax, Head-to-Head, Stress/Pressure Conditions
- Real-time Web UI: Live board animation, WebSocket updates, terminal-style dashboard
- Structured LLM Outputs: OpenAI JSON schema + Anthropic forced tool_use
- 5 Core Metrics: JSON compliance, move legality, win rate, avg latency, strategic quality
- 92 Unit Tests passing with full game engine, minimax, and orchestrator coverage
TreatyForge
Multi-agent LLM Diplomacy Benchmark — three AI agents compete in a turn-based territory game where they negotiate formal treaties, form alliances, and strategically betray commitments. Measures negotiation, deception, trust calibration, and the "say-do gap" across Claude models.
key_features
- 3-Player Diplomacy: 7-territory map, 12 turns, 9 treaty types (ceasefire, joint attack, conditional support, etc.)
- 14 LLM Variants + 5 Scripted Bots (Random, Greedy, Defensive, Treaty, Backstab)
- 5 Benchmark Phases: Treaty Comprehension, No Negotiation, Full Diplomacy, Stress, Adversarial Traps
- Say-Do Gap Analysis: Quantified divergence between diplomatic messages and actual military actions
- Chain-of-Thought Reasoning: Every action includes 2-3 sentences of strategic analysis
- Conversation Transcript Export: Full inter-model diplomatic messages as downloadable Markdown
- 109 Unit Tests passing across game engine, treaty engine, and integration suites
Atlas
Personal AI Terminal Assistant — a locally-running assistant powered by Gemma 4 E4B (4-bit, ~3 GB VRAM) with native tool calling. Chat naturally in the terminal to manage files, research the web, crawl sites, interact with APIs, work with git, and automate tasks. 71 tools, one consumer GPU, no cloud APIs.
key_features
- 71 Tools: Files, git, shell, web research, crawling, network, Docker, SQLite, PDF, images, clipboard, notes
- Deep Web Research: Search + parallel fetch top pages + LLM synthesis in one query
- Personal Knowledge Base: Atomic markdown notes with tags, search, and model-assisted recall
- 26 Deterministic Dispatch Patterns: Common queries bypass the LLM entirely in under 1 second
- Explain Before Execute: Risk-tiered plain-English summaries before every tool call
- Multi-round Tool Chaining: Up to 5 LLM rounds per query with retry, dedup, and hallucination correction
- 37 Tests passing with full tool, dispatch, explain, and web coverage
Professional Experience
Building AI solutions across industries.
AI Agent Engineer
AumeraAI · Solothurn, Switzerland
- AI Agent Reasoning
- Evaluations & Optimizations
AI/ML Engineer — Analytics
iProject LLC · Athens, Greece
- Established AI Department: built the AI Intelligence division from scratch
- Deployed various production AI applications including NLQ business intelligence tools
- Developed predictive analytics engines for F&B businesses
- Created AI lead generation systems and specialized chatbots
- Architected scalable AI solutions integrating multiple LLM providers
- Established AI governance and best practices organization-wide
Risk & Fraud Analyst
Kaizen Gaming · Athens, Greece
- Conducted advanced risk analysis and fraud pattern detection
- Maintained internal risk intelligence systems (Confluence documentation)
- Developed patterns to anticipate market trends and fraudulent behavior
- Designed fraud prevention controls reducing risk exposure
- Investigated complex fraud cases and established behavioral profiling frameworks
Mathematics & Physics Educator
Self-Employed · Athens, Greece
- Enhanced communication abilities through explaining advanced concepts to diverse audiences
- Developed teaching methodologies for complex mathematical and physical concepts
Technical Skills
Expertise in AI/ML, programming, and data analytics.
- LangGraph & Multi-Agent Systems
- TensorFlow
- Hugging Face Transformers
- Scikit-learn & NumPy
- RAG & Vector Databases (FAISS)
- Prompt Engineering & Fine-tuning
- Python (Advanced)
- R (Statistical Computing)
- SQL & NoSQL Databases
- FastAPI & Flask
- Git & Version Control
- REST APIs & WebSockets
- Microsoft Azure (AI Services)
- Docker & Containerization
- Model Deployment & Serving
- CI/CD Pipelines
- GPU Infrastructure Management
- API Development & Integration
- Statistical Analysis & Modeling
- Time-Series Forecasting
- Pandas, NumPy & SciPy
- Data Visualization (Plotly, Matplotlib)
- Business Intelligence
- A/B Testing & Experimentation
Education & Certifications
Academic background and professional certifications.
MSc Applied Mathematics & Statistics
National Technical University of Athens (NTUA)
School of Applied Mathematical and Physical Sciences (MSc Integrated Degree)
Thesis: Econometric Methods of Cryptocurrency Volatility Estimation (R Programming)
grade: 9.5 / 10
professional_certifications [13]
languages_and_soft_skills
Languages
- Greek (Native)
- Romanian (Native)
- English (C2 ECPE – Proficient)
Soft skills
- Critical Thinking
- Problem Solving
- Communication
- Team Collaboration
- Leadership
interests_and_passions
Let's Connect
Interested in AI/ML solutions or want to discuss potential collaborations? Feel free to reach out!
uplink: hello@ibanousis.tech