About · > szia.ai_

szia.ai explains how AI works and how to apply it effectively — a resource for leaders driving AI transformation, and a good read for anyone who’s simply curious.

I write this blog as a data science leader working on several generative AI projects. The focus is on building enterprise solutions that people actually use: GenAI where it shines, and traditional methods where control, precision, and auditability matter most.

Always open to grounded, genuinely useful conversations about applied GenAI — whether for internal initiatives or public settings: talks, panels, podcasts, conferences. That’s what the Talk & Connect page is for.

Marton Szel

Director of Data Science at Lynx · Singapore / HungaryDirector of Data Science at Lynx · SG / HU

X / TwitterX LinkedIn marton@szia.aiEmail

Project highlights · 2013 → 2026

2026 · Singapore

lynx-agents: An Orchestrator for Agentic Apps That Learn

Prototyped a typed orchestrator for agentic apps, a thin layer over Pydantic AI. Apps compile into a graph brain — scenarios, states, tools, worked example pathways — and run as state machines with guardrails enforced by construction, not by prompt. Pathways keep small LLMs on proven routes; smarter models are free to find better ones, and winning deviations are promoted into the brain — apps that learn from usage.

lynx-agents Brain Viewer: the compiled graph brain — scenarios, states, tools and example pathways — replaying a conversation

screenshot — lynx-agents Brain Viewer

2025/26 · Singapore

Agentic Applications for Interactive Analytics

Talk-to-Your-Data systems that let users query databases and documents in natural language — from raw stats to PowerPoint-ready summaries — with an agent guiding analysts end-to-end and suggesting next steps. Underneath: a knowledge graph with a novel retrieval method. Plus promptable feedback segmentation: describe how you want free-text comments split, get instant, refinable clustering.

m&m's in different colors — as an analogy to clustering

photo — Irfan Zaini

2024/25 · Singapore

LynxScribe: A Flexible Framework for Enterprise-Grade GenAI

A modular toolkit for building enterprise GenAI applications from reusable blocks — chatbots, RAG pipelines, agent workflows, orchestration layers — with full compatibility across clouds, services and on-prem LLMs, and fast LLM switching. A trainable Graph RAG memory underpins most applications for more reliable, context-aware outputs.

LynxScribe architecture — a modular agentic app builder platform from Lynx Analytics

diagram — LynxScribe modules

2023/24 · Singapore

Generative AI for Mathematical Theorem Proving

Reasoning systems that help LLMs navigate formalized proofs: step-by-step proof generation, heuristic guidance, search and reinforcement-learning loops, and synthetic training data. Exploring how generative AI supports symbolic reasoning — including uncovering novel problem-solving approaches and selecting Erdős-style theorems that open pathways for further work.

The graph of mathematics (Mathlib dependency graph in LEAN4)

graph — Mathlib imports

2023/24 · SingaporeNVIDIA GTC 2024

Graph RAG for Multi-Document Retrieval

A graph-based RAG where document chunks form nodes and trainable edges propagate similarity — retrieving all relevant documents with fewer queries. Deployed in copilot chatbots that learned from agent feedback, cutting average response times from six minutes to one with >80% of answers accepted without modification. The latest iteration adds ontology connections and LLM-generated artificial nodes.

Library — representing the knowledge base

photo — Ming Han Low

2020/22 · China & Global

Assortment Optimization for a Global Fashion Retailer

A framework maximizing store-level revenue within each market: historical and external data — weather patterns, luxury-brand clustering, city-edge outlet dynamics — shaped “store fingerprints” that guided assortment decisions and predicted sales even for products never stocked in a given store. Deployed first in China, then extended to US and EU markets.

Jeans in a fashion store, illustrating assortment optimization

photo — Martin Bammer

2018 · Germany

Fibre Network Extension Design with AlphaGo-Inspired Algorithms

Fibre roll-out formalized as a Prize-Collecting Steiner Tree on the road network — maximizing expected revenue from new household connections while minimizing deployment costs. Early Monte Carlo Tree Search experiments inspired by AlphaGo evolved into a dual ascent branch-and-bound method, shipped in LynxKite as a new graph optimization function.

photo — Elena Popova

2016/17 · Hong Kong

Clustering SIM Cards into Households Using Graphs

Graph algorithms building single-customer and household views for a telecom provider: positive and negative edges from shared addresses, common night locations, co-location and calling patterns — while excluding impossible connections. An optimization pass balanced weak and strong signals into the most probable groupings, from individuals to multi-generational households.

photo — Chi Hung Wong

2013/14 · Hungary

Graph Models for VAT Fraud Detection

Graph models uncovering carousel VAT fraud — chains of fictitious companies issuing and reclaiming VAT with no real economic activity — and estimating the illicit gains of hidden beneficiaries from real invoice data. In parallel: early-warning models on a firm-level knowledge graph, flagging high-risk companies at foundation time via ownership links and shared addresses.

photo — Alina Grubnyak