talk-data.com

​Alexey Grigorev, the course creator, will officially start the new cohort of the Data Engineering Zoomcamp in this live session. He’ll walk you through the course structure, key topics, and what you’ll build.

​What You’ll Learn During the Session

​Alexey will walk you through:

• ​What’s included in the course: topics, tools, and hands-on projects
• ​How assignments, feedback, and scoring work
• ​How to approach the material, even if you’re new to ML
• ​What it’s like to learn together with thousands of learners in the DataTalks.Club community

​You’ll also have a chance to ask Alexey your questions live.

​Thinking About AI Dev Tools Zoomcamp?

​Data Engineering Zoomcamp is a free 9-week course covering infrastructure setup, workflow orchestration, data warehousing, analytics, batch processing, and streaming. The last three weeks focus on a capstone project in which you'll build an end-to-end data pipeline using a dataset of your choice, demonstrating data lake and warehouse solutions with documentation. Projects are peer-reviewed by fellow participants. ​The new cohort of the Data Engineering Zoomcamp starts on January 12, 2026. You can join by registering here. ​

​About the Speaker

​Alexey Grigorev is the Founder of DataTalks.Club and creator of the Zoomcamp series. ​Alexey is a seasoned software and ML engineer with over 10 years of engineering experience and 6+ years in machine learning. He has deployed large-scale ML systems at companies like OLX Group and Simplaex, authored several technical books, including Machine Learning Bookcamp, and is a Kaggle Master with a 1st place finish in the NIPS'17 Criteo Challenge. ​

Join our slack: https://datatalks.club/slack.html

Build a Multi-Agent Deep Research System with Temporal - Alexey Grigorev

​In this hands-on workshop, you'll build a durable deep-research agent and learn how to make LLM-powered systems reliable enough for real production environments.

​We’ll walk through:

• ​Building a deep-research agent step by step
• ​Handling failures and unpredictable LLM behavior
• ​Designing agentic workflows with Temporal + PydanticAI

​By the end of the workshop, you'll know how to take an idea from PoC to a production-grade multi-agent system with Temporal: observable, fault-tolerant, easy to extend, and designed to survive real-life conditions. ​ ​About the speaker:

​Alexey Grigorev is the Founder of DataTalks.Club and creator of the Zoomcamp series.

​Alexey is a seasoned software and ML engineer with over 10 years in engineering and 6+ years in machine learning. He has deployed large-scale ML systems at companies like OLX Group and Simplaex, authored several technical books including Machine Learning Bookcamp, and is a Kaggle Master with a 1st place finish in the NIPS'17 Criteo Challenge. ​

​​​​​​​DataTalks.Club is the place to talk about data. Join our slack community! ​

​This event is sponsored by Temporal.

Curious about the Data Engineering Zoomcamp? Join us for a live, interactive Q&A session with course creator Alexey Grigorev and get all your questions answered before the new cohort begins on January 12, 2026.

​What You’ll Learn During the Session

​Alexey will walk you through:

• ​Course curriculum: An overview of the topics, tools, and real-world projects you'll work on
• ​Assignments and grading: How homework and peer reviews are structured
• ​Getting help and feedback: How support works during the course and in the community
• ​Portfolio-building projects: Which hands-on projects will you build during the course to showcase your skills
• ​Learning with the community: How to engage with 79k+ learners in the DataTalks.Club Slack

​Alexey will also share tips on how to follow the material effectively, pace your learning, and stay motivated throughout the course. ​You’ll get a chance to meet the instructor, learn more about the course structure, and ask your questions. ​

​Thinking About Data Engineering Zoomcamp?

​Data Engineering Zoomcamp is a free 9-week course covering infrastructure setup, workflow orchestration, data warehousing, analytics, batch processing, and streaming. The last three weeks focus on a capstone project in which you'll build an end-to-end data pipeline using a dataset of your choice, demonstrating data lake and warehouse solutions with documentation. Projects are peer-reviewed by fellow participants. ​The new cohort of the Data Engineering Zoomcamp starts on January 12, 2026. You can join by registering here. ​

​About the Speaker

​Alexey Grigorev is the Founder of DataTalks.Club and creator of the Zoomcamp series. ​Alexey is a seasoned software and ML engineer with over 10 years in engineering and 6+ years in machine learning. He has deployed large-scale ML systems at companies like OLX Group and Simplaex, authored several technical books including Machine Learning Bookcamp, and is a Kaggle Master with a 1st place finish in the NIPS'17 Criteo Challenge.

Join our slack: https://datatalks.club/slack.html

Alexey Grigorev is hosting a live hands-on workshop to explore how Docker can simplify your data workflows, from setting up databases to packaging your scripts for reproducibility. ​This session is open to everyone interested in learning practical Docker skills for data engineering and analytics. ​During the workshop, we’ll walk through a complete workflow using Docker, PostgreSQL, pgAdmin, and Docker Compose, showing how to run and connect multiple services with minimal setup effort. ​ ​What you’ll learn:

• ​Why Docker is essential for modern data work
• ​Running PostgreSQL in a Docker container and connecting it to your local environment
• ​Connecting with pgcli or directly from Jupyter notebooks
• ​Loading and exploring the NYC Taxi dataset
• ​Managing databases visually with pgAdmin 4 (latest UI)
• ​Converting notebooks to scripts and Dockerizing them
• ​Running multi-container setups with Docker Compose
• ​Understanding Docker networking, port mapping, and volumes

​The workshop will be recorded and later used to refresh the Docker module of the Data Engineering Zoomcamp, so you’ll also get a preview of what’s coming in the new course release. ​ ​Thinking about DE Zoomcamp? ​Data Engineering Zoomcamp is a free 9-week course on building production-ready data pipelines. The next cohort starts in January 2026. Join the course here. ​ ​About the Speaker ​Alexey Grigorev is the Founder of DataTalks.Club and creator of the Zoomcamp series. ​​Alexey is a seasoned software and ML engineer with over 10 years in engineering and 6+ years in machine learning. He has deployed large-scale ML systems at companies like OLX Group and Simplaex, authored several technical books including Machine Learning Bookcamp, and is a Kaggle Master with a 1st place finish in the NIPS'17 Criteo Challenge.

Join our slack: https://datatalks.club/slack.html ​

In this talk, Anusha Akkina, co-founder of Auralytix, shares her journey from working as a Chartered Accountant and Auditor at Deloitte to building an AI-powered finance intelligence platform designed to augment, not replace, human decision-making. Together with host Alexey from DataTalks.Club, she explores how AI is transforming finance operations beyond spreadsheets—from tackling ERP limitations to creating real-time insights that drive strategic business outcomes.

TIMECODES: 00:00 Building trust in AI finance and introducing Auralytix 02:22 From accounting roots to auditing at Deloitte and Paraxel 08:20 Moving to Germany and pivoting into corporate finance 11:50 The data struggle in strategic finance and the need for change 13:23 How Auralytix was born: bridging AI and financial compliance 17:15 Why ERP systems fail finance teams and how spreadsheets fill the gap 24:31 The real cost of ERP rigidity and lessons from failed transformations 29:10 The hidden risks of spreadsheet dependency and knowledge loss 37:30 Experimenting with ChatGPT and coding the first AI finance prototype 43:34 Identifying finance’s biggest pain points through user research 47:24 Empowering finance teams with AI-driven, real-time decision insights 50:59 Developing an entrepreneurial mindset through strategy and learning 54:31 Essential resources and finding the right AI co-founder

Connect with Anusha - Linkedin - https://www.linkedin.com/in/anusha-akkina-acma-cgma-56154547/ - Website - https://aurelytix.com/

Connect with DataTalks.Club: - Join the community - https://datatalks.club/slack.html - Subscribe to our Google calendar to have all our events in your calendar - https://calendar.google.com/calendar/r?cid=ZjhxaWRqbnEwamhzY3A4ODA5azFlZ2hzNjBAZ3JvdXAuY2FsZW5kYXIuZ29vZ2xlLmNvbQ - Check other upcoming events - https://lu.ma/dtc-events - GitHub: https://github.com/DataTalksClub - LinkedIn - https://www.linkedin.com/company/datatalks-club/ - Twitter - https://twitter.com/DataTalksClub - Website - https://datatalks.club/

Neural Radiance Field (NeRF)-based segmentation methods focus on object semantics and rely solely on RGB data, lacking intrinsic material properties. We introduce UnMix-NeRF, a framework that integrates spectral unmixing into NeRF, enabling joint hyperspectral novel view synthesis and unsupervised material segmentation. Our method models spectral reflectance via diffuse and specular components, where a learned dictionary of global endmembers represents pure material signatures, and per-point abundances capture their distribution. For material segmentation, we use spectral signature predictions along learned endmembers, allowing unsupervised material clustering. Additionally, UnMix-NeRF enables scene editing by modifying learned endmember dictionaries for flexible material-based appearance manipulation. Extensive experiments validate our approach, demonstrating superior spectral reconstruction and material segmentation to existing methods.

We tackle the challenge of jointly personalizing content and style from a few examples. A promising approach is to train separate Low-Rank Adapters (LoRA) and merge them effectively, preserving both content and style. Existing methods, such as ZipLoRA, treat content and style as independent entities, merging them by learning masks in LoRA's output dimensions. However, content and style are intertwined, not independent. To address this, we propose DuoLoRA, a content-style personalization framework featuring three key components: (i) rank-dimension mask learning, (ii) effective merging via layer priors, and (iii) Constyle loss, which leverages cycle-consistency in the merging process. First, we introduce ZipRank, which performs content-style merging within the rank dimension, offering adaptive rank flexibility and significantly reducing the number of learnable parameters. Additionally, we incorporate SDXL layer priors to apply implicit rank constraints informed by each layer's content-style bias and adaptive merger initialization, enhancing the integration of content and style. To further refine the merging process, we introduce Constyle loss, which leverages the cycle-consistency between content and style. Our experimental results demonstrate that DuoLoRA outperforms state-of-the-art content-style merging methods across multiple benchmarks.

Are Vision-Language Models Ready for Physical AI? Humans easily understand how objects move, rotate, and shift while current AI models that connect vision and language still make mistakes in what seem like simple situations: deciding “left” versus “right” when something is moving, recognizing how perspective changes, or keeping track of motion over time. To reveal these kinds of limitations, we created VLM4D, a testing suite made up of real-world and synthetic videos, each paired with questions about motion, rotation, perspective, and continuity. When we put modern vision-language models through these challenges, they performed far below human levels, especially when visual cues must be combined or the sequence of events must be maintained. But there is hope: new methods such as reconstructing visual features in 4D and fine-tuning focused on space and time show noticeable improvement, bringing us closer to AI that truly understands a dynamic physical world.

CLIP is a foundational model with transferable classification performance in the few-shot setting. Several methods have shown improved performance of CLIP using few-shot examples. However, so far, all these techniques have been benchmarked using standard few-shot datasets. We argue that this mode of evaluation does not provide a true indication of the inductive generalization ability using few-shot examples. As most datasets have been seen by the CLIP model, the resultant setting can be termed as partially transductive. To solve this, we propose a pipeline that uses an unlearning technique to obtain true inductive baselines. In this new inductive setting, the methods show a significant drop in performance (-55% on average among 13 baselines with multiple datasets). We validate the unlearning technique using oracle baselines. An improved few-shot classification technique is proposed that consistently obtains state-of-the-art performance over 13 other recent baseline methods on a comprehensive analysis with 5880 experiments - varying the datasets, differing number of few-shot examples, unlearning setting, and with different seeds. Thus, we identify the issue with the evaluation of CLIP-based few-shot classification, provide a solution using unlearning, propose new benchmarks, and provide an improved method.

CLIP is a foundational model with transferable classification performance in the few-shot setting. Several methods have shown improved performance of CLIP using few-shot examples. However, so far, all these techniques have been benchmarked using standard few-shot datasets. We argue that this mode of evaluation does not provide a true indication of the inductive generalization ability using few-shot examples. As most datasets have been seen by the CLIP model, the resultant setting can be termed as partially transductive. To solve this, we propose a pipeline that uses an unlearning technique to obtain true inductive baselines. In this new inductive setting, the methods show a significant drop in performance (-55% on average among 13 baselines with multiple datasets). We validate the unlearning technique using oracle baselines. An improved few-shot classification technique is proposed that consistently obtains state-of-the-art performance over 13 other recent baseline methods on a comprehensive analysis with 5880 experiments - varying the datasets, differing number of few-shot examples, unlearning setting, and with different seeds. Thus, we identify the issue with the evaluation of CLIP-based few-shot classification, provide a solution using unlearning, propose new benchmarks, and provide an improved method.

Tracking and forecasting the rotation of objects is fundamental in computer vision and robotics, yet SO(3) extrapolation remains challenging as (1) sensor observations can be noisy and sparse, (2) motion patterns can be governed by complex dynamics, and (3) application settings can demand long-term forecasting. This work proposes modeling continuous-time rotational object dynamics on SO(3) using Neural Controlled Differential Equations guided by Savitzky-Golay paths. Unlike existing methods that rely on simplified motion assumptions, our method learns a general latent dynamical system of the underlying object trajectory while respecting the geometric structure of rotations. Experimental results on real-world data demonstrate compelling forecasting capabilities compared to existing approaches.

Are Vision-Language Models Ready for Physical AI? Humans easily understand how objects move, rotate, and shift while current AI models that connect vision and language still make mistakes in what seem like simple situations: deciding left versus right when something is moving, recognizing how perspective changes, or keeping track of motion over time. To reveal these kinds of limitations, we created VLM4D, a testing suite made up of real-world and synthetic videos, each paired with questions about motion, rotation, perspective, and continuity. When we put modern vision-language models through these challenges, they performed far below human levels, especially when visual cues must be combined or the sequence of events must be maintained. But there is hope: new methods such as reconstructing visual features in 4D and fine-tuning focused on space and time show noticeable improvement, bringing us closer to AI that truly understands a dynamic physical world.

Neural Radiance Field (NeRF)-based segmentation methods focus on object semantics and rely solely on RGB data, lacking intrinsic material properties. We introduce UnMix-NeRF, a framework that integrates spectral unmixing into NeRF, enabling joint hyperspectral novel view synthesis and unsupervised material segmentation. Our method models spectral reflectance via diffuse and specular components, where a learned dictionary of global endmembers represents pure material signatures, and per-point abundances capture their distribution. For material segmentation, we use spectral signature predictions along learned endmembers, allowing unsupervised material clustering. Additionally, UnMix-NeRF enables scene editing by modifying learned endmember dictionaries for flexible material-based appearance manipulation. Extensive experiments validate our approach, demonstrating superior spectral reconstruction and material segmentation to existing methods.

Neural Radiance Field (NeRF)-based segmentation methods focus on object semantics and rely solely on RGB data, lacking intrinsic material properties. We introduce UnMix-NeRF, a framework that integrates spectral unmixing into NeRF, enabling joint hyperspectral novel view synthesis and unsupervised material segmentation. Our method models spectral reflectance via diffuse and specular components, where a learned dictionary of global endmembers represents pure material signatures, and per-point abundances capture their distribution. For material segmentation, we use spectral signature predictions along learned endmembers, allowing unsupervised material clustering. Additionally, UnMix-NeRF enables scene editing by modifying learned endmember dictionaries for flexible material-based appearance manipulation. Extensive experiments validate our approach, demonstrating superior spectral reconstruction and material segmentation to existing methods.

We tackle the challenge of jointly personalizing content and style from a few examples. A promising approach is to train separate Low-Rank Adapters (LoRA) and merge them effectively, preserving both content and style. Existing methods, such as ZipLoRA, treat content and style as independent entities, merging them by learning masks in LoRA's output dimensions. However, content and style are intertwined, not independent. To address this, we propose DuoLoRA, a content-style personalization framework featuring three key components: (i) rank-dimension mask learning, (ii) effective merging via layer priors, and (iii) Constyle loss, which leverages cycle-consistency in the merging process. First, we introduce ZipRank, which performs content-style merging within the rank dimension, offering adaptive rank flexibility and significantly reducing the number of learnable parameters. Additionally, we incorporate SDXL layer priors to apply implicit rank constraints informed by each layer's content-style bias and adaptive merger initialization, enhancing the integration of content and style. To further refine the merging process, we introduce Constyle loss, which leverages the cycle-consistency between content and style. Our experimental results demonstrate that DuoLoRA outperforms state-of-the-art content-style merging methods across multiple benchmarks.

Tracking and forecasting the rotation of objects is fundamental in computer vision and robotics, yet SO(3) extrapolation remains challenging as (1) sensor observations can be noisy and sparse, (2) motion patterns can be governed by complex dynamics, and (3) application settings can demand long-term forecasting. This work proposes modeling continuous-time rotational object dynamics on SO(3) using Neural Controlled Differential Equations guided by Savitzky-Golay paths. Unlike existing methods that rely on simplified motion assumptions, our method learns a general latent dynamical system of the underlying object trajectory while respecting the geometric structure of rotations. Experimental results on real-world data demonstrate compelling forecasting capabilities compared to existing approaches.

Are Vision-Language Models Ready for Physical AI? Humans easily understand how objects move, rotate, and shift while current AI models that connect vision and language still make mistakes in what seem like simple situations: deciding “left” versus “right” when something is moving, recognizing how perspective changes, or keeping track of motion over time. To reveal these kinds of limitations, we created VLM4D, a testing suite made up of real-world and synthetic videos, each paired with questions about motion, rotation, perspective, and continuity. When we put modern vision-language models through these challenges, they performed far below human levels, especially when visual cues must be combined or the sequence of events must be maintained. But there is hope: new methods such as reconstructing visual features in 4D and fine-tuning focused on space and time show noticeable improvement, bringing us closer to AI that truly understands a dynamic physical world.

We tackle the challenge of jointly personalizing content and style from a few examples. A promising approach is to train separate Low-Rank Adapters (LoRA) and merge them effectively, preserving both content and style. Existing methods, such as ZipLoRA, treat content and style as independent entities, merging them by learning masks in LoRA's output dimensions. However, content and style are intertwined, not independent. To address this, we propose DuoLoRA, a content-style personalization framework featuring three key components: (i) rank-dimension mask learning, (ii) effective merging via layer priors, and (iii) Constyle loss, which leverages cycle-consistency in the merging process. First, we introduce ZipRank, which performs content-style merging within the rank dimension, offering adaptive rank flexibility and significantly reducing the number of learnable parameters. Additionally, we incorporate SDXL layer priors to apply implicit rank constraints informed by each layer's content-style bias and adaptive merger initialization, enhancing the integration of content and style. To further refine the merging process, we introduce Constyle loss, which leverages the cycle-consistency between content and style. Our experimental results demonstrate that DuoLoRA outperforms state-of-the-art content-style merging methods across multiple benchmarks.