Looking for a job

Alex Fafula

@alexfafula

AI inference engineer at Biuro Informacji Kredytowej, serving a 135 GB model on NVIDIA DGX Spark with 4× smaller embedding memory.

Poland

Message

At Biuro Informacji Kredytowej, I build low-latency inference pipelines for private open-source models in an air-gapped banking network. I also designed regression tests for structured entity extraction and model output stability.

On OneSpark, I served a 135 GB model on NVIDIA DGX Spark under SGLang. I used HashK compression to reduce the embedding table from 51.2 GB to 12.8 GB, with no degradation on complex reasoning benchmarks.

At Chabre Data & AI Advisory, I advised technology clients on moving from commercial LLM APIs to self-hosted inference, and benchmarked runtimes to reduce cloud compute costs by up to 65%. I also published research on model refusal removal and its effects on decision disposition.

Experience

Work history, roles, and key accomplishments

Current

Lead AI Architect

Current

Biuro Informacji Kredytowej

Jul 2025 - Present (1 year 2 months)

Directed machine learning systems and GPU infrastructure for Poland's statutory credit bureau, managing on-premises hardware clusters and model serving runtimes. Built low-latency inference pipelines running private open-source models in an isolated, air-gapped banking network.

GPU Infrastructure Model Serving Air Gapped Systems Low Latency Inference Serving Open Source Models Regression Testing Evaluation Suites EBA GDPR DORA

Chief Data Officer / Principal Systems Architect

Chabre Data & AI Advisory

Jan 2023 - Dec 2025 (2 years 11 months)

Advised 15+ international technology clients on migrating from commercial LLM APIs to self-hosted open-source inference stacks on private cloud and bare-metal GPU infrastructure. Benchmarked and optimized inference runtimes across various hardware configurations, achieving up to 65% reduction in cloud compute costs.

VLLM Sglang NVIDIA TensorRT LLM TGI AWQ GPTQ FP8 KV Cache Management Prefix Caching Model Evaluation Harnesses Model Compression

Director, Data Architecture & Analytics

Diversey

Jan 2023 - Dec 2023 (11 months)

Led edge telemetry and low-level software engineering across Diversey's global fleet of 1,200 autonomous industrial cleaning robots. Deployed real-time anomaly detection models directly on edge microcontrollers and embedded Linux gateways.

Edge Computing Embedded Linux Anomaly Detection MQTT Kafka Sensor data processing C C++

Global Enterprise Solutions Architect

British American Tobacco

Jan 2020 - Dec 2023 (3 years 11 months)

Governed reference architectures for enterprise data platforms, machine learning workloads, and cloud environments across 90,000+ employees in 180 countries. Directed modernization of high-throughput analytics platforms on Databricks and cloud environments.

Enterprise Architecture Databricks Azure AWS Microservices GPU Accelerated Computing Performance Benchmarking Cost Governance

Assistant Professor & Lecturer

Wrocław University of Economics

Jan 2005 - Dec 2022 (17 years 11 months)

Taught post-graduate and MBA courses covering High-Performance Computing, Compiler Design, Machine Learning Architectures, and Applied Econometrics.

High Performance Computing Compiler Design Principles Machine Learning Econometrics Teaching

Head of AI & Data Science Infrastructure

Publicis Groupe

Jan 2018 - Dec 2020 (2 years 11 months)

Architected real-time bidding algorithms and programmatic audience segmentation systems processing billions of events daily with sub-50ms latency SLAs. Built distributed Kubernetes clusters and model serving infrastructure with automated health checks and traffic shifting.

Real Time Bidding Programmatic Advertising Kubernetes Model Serving Distributed Systems Low Latency Systems

Data Analytics & Big Data, Manager

PwC Poland

Jan 2015 - Dec 2018 (3 years 11 months)

Architected high-throughput big data processing engines and real-time fraud detection systems on Cloudera Hadoop for tier-1 commercial banks and telecom operators. Built M.A.I.A., an interactive NLP assistant for automated profile evaluation and recruitment screening.

Big Data Hadoop Cloudera Fraud Detection NLP Distributed Systems Streaming Data

Chief Data Scientist

MarketPsych

Jan 2011 - Dec 2016 (5 years 11 months)

Architected high-throughput ingestion and NLP pipelines analyzing millions of global news and social records daily for the Thomson Reuters / LSEG MarketPsych Indices used by Wall Street quant funds. Designed low-latency string parsing, sentiment extraction, and market entity disambiguation algorithms.

NLP Sentiment Analysis High Volume Ingestion Low Latency Parsing Disambiguation Financial Data

Education

Degrees, certifications, and relevant coursework

Wrocław University of Economics

PhD with Distinction, Machine Learning Applied to Financial Decision-Making

2007 - 2014

Doctoral dissertation awarded with honours; computational modeling of decision-making under uncertainty. Formally recognized with an academic award by the Scientific Society for Economic Informatics (NTIE).