Feliks Bańka
AI Engineer & Research Engineer
↓
AI Engineer & Research Engineer
I am a master’s student in Artificial Intelligence at Warsaw University of Technology and an AI Engineer at AI Clearing, where I build production LLM systems, RAG pipelines, and multi-agent architectures for complex analytical workflows. My research focuses on LLM evaluation, evaluator adaptation, multi-agent systems, cooperative AI, and reliable AI-generated outputs. Recent work includes sequential evaluation of models adapting to hidden evaluator policies, bounded verification of AI-generated knowledge artifacts, and cooperative multi-agent reinforcement learning. I have authored and co-authored work accepted at venues including NeurIPS 2026, ICADL 2026, ICCS 2026, PACIS 2025, ACIIDS 2025, and AAMAS workshops, with current work under review or submitted to ICLR 2027, AAAI 2027, and AAMAS 2027.
Major: Artificial Intelligence • GPA: 4.60/5.0
Thesis: Investigating the Application of Game Theory in LLM-based Multi-Agent Systems
Major: Automatic Control & Robotics
Thesis: AI Architectures for Option Pricing and Portfolio Management: Transformers to MAS
Built and ran a controlled sequential LLM evaluation environment testing adaptation to hidden evaluator policies through repeated feedback, including policy-switch, information-seeking, and matched-control experiments across multiple model families.
Studies how structured critic designs can represent both shared team-level value and agent-specific disagreement in cooperative policy-gradient learning.
Studies bounded audits of LLM outputs, including claim composition, generator-reviewer incentives, and informed inspection of decision-critical claims.
Proposes a PPO-compatible critic design with a shared team baseline and zero-mean agent-specific residuals to improve credit assignment in cooperative multi-agent reinforcement learning.
Examines how verification capacity and reliability change when AI-generated artifacts contain more claims than a bounded reviewer can inspect.
Introduces a retrieval-augmented multi-agent framework for portfolio decision support, combining historical precedent retrieval, structured allocation decisions, and risk-aware evaluation.
Studies explanation-aware multi-agent coordination under asymmetric information and models LLM communication as a signaling problem with verifiable reasoning.
Presents a multi-agent reinforcement learning framework for portfolio optimization and options-based hedging under changing market conditions.
Integrated neural networks, volatility modeling, sentiment analysis via LLMs, and reinforcement learning into a hybrid option-pricing and trading platform.
Compares Transformer-based forecasting architectures with classical and deep sequence baselines across equities, indices, and crypto options.
Adapts the Informer architecture to option pricing and studies long-sequence modeling for derivatives forecasting.
Architected and deployed a production multi-agent LLM system using LangGraph and LangChain, including stateful workflows, tool routing, backend API calls, and RAG. Built async FastAPI services with strict Pydantic v2 schemas for reliable structured outputs and downstream integrations, and designed retrieval pipelines over heterogeneous enterprise data using OCR, vision LLMs, BM25, reranking, vector search, and selective context injection.
Impact: Increased benchmark accuracy from 78% to 98% through evaluation, failure analysis, and regression testing. Reduced latency from 80s to 20s and token usage from 1.1M to 200k through caching, selective routing, and context optimization.
Developed Python-based automation tools and data pipelines. Analyzed market trends and optimized strategies to support business operations.
Built React (TypeScript) and .NET micro-frontends, implemented REST APIs, and integrated backend logic in cross-functional Agile teams.
Built a controlled sequential evaluation environment for studying how LLMs adapt to hidden evaluator policies through repeated PASS/FAIL feedback, including policy-switch, information-seeking, matched-control, and multi-model experiments.
Built a research-oriented market-making simulation framework based on inventory-aware quoting, partial fills, fees/slippage, risk controls, calibration utilities, and reproducible strategy benchmarking.
Developed an event-level order-book research pipeline for short-horizon mid-price prediction under latency, transaction cost, and anti-leakage constraints, with causal features, chronological validation, and post-cost trading diagnostics.
Leading a student research group focused on generative AI, multi-agent systems, LLM applications, and argument modeling. Promoting practical AI applications in finance and decision systems.
Led AI/LLM-oriented student projects, organized tech-business meetups, and managed a 10-member team delivering workshops and hackathons in Warsaw.
Collaborated on entrepreneurship projects addressing social challenges through innovation, with emphasis on leadership and cross-functional teamwork.
Top-ranked student paper by the program committee.
Individual research funding.
National award granted to 400 top students nationwide.