DataFest — Belgrade, May 31

FON, University of Belgrade

---

12:00 — Guest Registration & Welcome

12:45 — Opening Remarks · Salavat Gariffullin, ODS Serbia

12:50 — Welcome Address

Prof. Dragan Vukmirović, PhD (FON University)
From Big Data to AI-native: 7V, Synthetic Data and the New Role of Data Science in Industry

13:00 — Stage 1: Agents & LLMs · Stage 2: RecSys ML

14:40 — Coffee Break

15:10 — Stage 1: Robotics & CV · Stage 2: Banking Language Models

16:25 — Lunch Break

17:25 — Stage 1: Data Quality · Stage 2: Voice ML

19:05 — Closing Remarks · Salavat Gariffullin, ODS Serbia

19:15 — After-party 🎉

---

---

Networking Lounge · Library

Open for networking throughout the event, especially during breaks.

---

---

🏛 Stage 1 · Amphitheater 1

Agents & LLMs

Time	Speaker	Company	Talk	Description
13:00	Ivan Bushmarinov	Perplexity	User-Guided LLM Answer Quality Evaluation	Leveraging thread-style user feedback and small trained models to evaluate frontier LLM answers and enable scalable benchmarking.
13:25	Ksenija Blažević	Lemana Tech	How not to build agentic AI: 4 (very common) anti-patterns and what to do instead	Common anti-patterns in agentic AI and how to replace them with leaner, cost-efficient architectures.
13:50	Dmitrii Krasnov	Zencoder	Orchestrating Coding Models	Comparison of sequential, parallel, and OSS orchestration for coding models and their impact on SWE-bench-like benchmarks.
14:15	Michael Diskin	HSE University	When Models Should Stay Silent	Measuring model uncertainty, calibrating confidence, and implementing rejection mechanisms for more reliable LLM systems in production.

14:40–15:10 — Coffee Break

Robotics & CV

Time	Speaker	Company	Talk	Description
15:10	Fedor Kurdov	Yandex	RL for Real-World Robot Motion Planning	How RL (without imitation learning) was built from scratch and deployed for motion planning in Yandex's sidewalk delivery rovers.
15:35	Dmitrii Iunovidov	LogicYield	Making Industrial CV Fly on Edge CPUs: A Neuro-Symbolic Benchmark for Dense Instance Segmentation	Running industrial computer vision on edge CPUs in harsh factory conditions using inference optimization and neuro-symbolic methods.
16:00	Aleksey Postnikov	Sber Robotics Lab	Physical AI: Status and the Road Ahead	Broad overview of Physical AI: synthetic data, sim-to-real, RL over behavior cloning, and learning policies from human videos.

16:25–17:25 — Lunch Break

Data Quality

Time	Speaker	Company	Talk	Description
17:25	Oleg Sekachev	Yandex	Agent for Data Labelling. LLM with Hammer and Ruler	Quality data labelling — faster and cheaper than humans, simpler and more accurate than a bare LLM.
17:50	Anastasiia Margolina	Banco Plata	How we (didn't) build an AutoEval	A story about evaluating AI when the answers are about real money — and how the autoeval we thought would be a single prompt turned into a methodology.
18:15	Stefan Hačko	Foursquare	LLM-Powered Harmonization of 100M+ Places	How Foursquare uses LLMs and vector embeddings to clean, match, and unify massive third-party venue datasets at scale.
18:40	Alexey Korotkov & Timofey Garaev	MIPT AI Institute	SHARP: Span-level Hallucination Annotation for Reasoning Paths	New span-level dataset for hallucination detection in LLM reasoning paths and why it yields better downstream quality for PRM models.

---

---

🏛 Stage 2 · Amphitheater 2

RecSys ML

Time	Speaker	Company	Talk	Description
13:00	Alexander Eroshenko	Yandex	LLM-Powered Item-to-Item Recs in Lavka	Practical case of deploying a compact LLM (Gemma ~270M) for item-to-item recommendations of substitutes and complements.
13:25	Vladimir Kukushkin	Independent Researcher	Beyond Funnels: Advanced UX Analytics	How to study user behavior deeper than traditional funnels using advanced UX analytics tools and user journey analysis.
13:50	Alexey Vasilev	Sber AI Lab	SplitLight: RecSys Evaluation Toolkit	Open-source toolkit for analyzing datasets and split strategies in RecSys to make offline evaluation transparent and reproducible.
14:15	Nikita Severin	Independent Researcher	Knowledge Transfer from Pre-trained LLMs to Recommender Models	Efficient knowledge transfer from pre-trained LLMs to recommender models without costly serving-time inference or architectural changes.

14:40–15:10 — Coffee Break

Banking Language Models

Time	Speaker	Company	Talk	Description
15:10	Boris Tseitlin	Banco Plata	Learning from Unstructured Sequences in 2026	Overview of self-supervised and foundation-model approaches to embeddings from transactions, events, and other unstructured sequences.
15:35	Mikhail Sysoev	Banco Plata	PV Models in Retail Lending	PV models and approaches to optimizing product parameters in card-based fintech products.
16:00	Victor Barbarich	Banco Plata	Transformers Replace Feature Engineering in Scoring	Moving from manual feature engineering in credit scoring to transformers that learn directly from raw account and employment histories.

16:25–17:25 — Lunch Break

Voice ML

Time	Speaker	Company	Talk	Description
17:25	Ilya Shigabeev	Langswap.app	AI video dubbing with open source — How we’ve built speech-to-speech pipeline and what we’ve learned from it	Challenges in AI dubbing: different text lengths after translation, preserving pauses from the source video, accent issues, and dependency hell — with a walkthrough of the open-source pipeline.
17:50	Pavel Mazaev	Yandex	Device-Directed Speech Detection for Alice	Production system for detecting speech directed at a smart device (Yandex Alice) to enable natural dialogue without a constant wake word.
18:15	Fedor Konovalenko	MIL Team (MIPT)	From Model Compression to Local Inference Platform	How a model compression tool evolved into a local GenAI inference platform with OpenAI-compatible API, multi-engine support, and observability.
18:40	Pavel Guliaev	Independent Researcher	Video2Text: Industry State & Practical Choices	State of the Video2Text industry — what works, current limitations, and how to pick and adapt solutions for production load and budget.