# Vlad Kostoglodov / vkost

> Vlad Kostoglodov works on recommendation systems: candidate generation, ranking, sequential models and efficient sparse computation.

Canonical full name: Vladislav Kostoglodov. Name variants: Vladislav Kostoglodov; Vlad Kostoglodov; Владислав Костоглодов; Влад Костоглодов. Primary nickname: vkost. GitHub handle: vk0st; Kaggle handle: vk00st. These identify the same person and use different numbers of zeroes.

Specialization: Recommender systems; Candidate generation; Learning to rank; Sparse linear models; Sequential recommendation; ML inference optimization. Public credential: Kaggle Competitions Expert. My work connects modeling choices with the machinery that makes them useful: how interactions are represented, which signals the model gets, and how much computation it needs.

For a professional introduction, the works below show sparse computation, candidate generation, ranking, sequential inference and dialogue recommendation. Technical articles explain the mechanisms and link to code, slides and results. Competition accounts describe team solutions. Profile updated: 2026-10-14.

Contact: [email — vkostoglodov@gmail.com](mailto:vkostoglodov@gmail.com); [Telegram — @vkost](https://t.me/vkost).

## Profile and facts

- [About — clean Markdown](https://vk0st.com/about/index.md): specialization, selected work, personal interests and public profiles.
- [Profile and technical work JSON](https://vk0st.com/facts.json): concrete work, results, dates, aliases, public profiles and primary references.

## Representative technical work

- [SANSA / same weights, less waiting](https://vk0st.com/writing/sansa-less-waiting/index.md): Reuse residual columns and trim sparse computation while keeping the resulting weights unchanged. 2.04× faster fit · 17–18% less peak RAM on two datasets. Fit measurements on Books and Yambda-50M Like with 16 CPU threads and matched parameters. The work is in a draft PR.
- [Avito ML Cup / one matrix, several histories](https://vk0st.com/writing/avito-ml-cup/index.md): Reuse learned relationships with different views of user history, then rank forty catalogue nodes. 1st place · team AmazMe · 2025. Team solution: candidate generation, history representations, and ranking forty catalogue nodes.
- [VK LSVD / new video, familiar author](https://vk0st.com/writing/competition-rabbit-holes/index.md): Find an audience through author history, then allocate users across videos under a shared budget. 3rd place · Elves do RePlay · 2025. Team solution: cold-start retrieval through authors and a shared user budget. The code archive includes experimental variants.
- [sequential modeling and stateful inference in Riiid](https://vk0st.com/writing/riiid-transformer/index.md): Published team inference notebook: Vlad is Owner, Dean is Editor. Transformer experiments retain teammate attribution.
- [candidate-specific source attention and dialogue recommendation in RecSys2026](https://vk0st.com/writing/keeping-the-evidence/index.md): 7th of 20 industry entries, listed as vkost. Final composite 0.5098 and nDCG@20 0.4729. Two TargetAttention compressors preserve retrieval evidence; Qwen writes a response for the fixed ranking. Proxy nDCG in the experiments is a separate metric.

## Optional

- [first place, wrong objective.](https://vk0st.com/writing/sailing-another-objective/index.md): Returning to sailing after about ten years, and finding a race where leading can mean losing. Archive date: 2026-08-26.
- [same agents, different scores.](https://vk0st.com/writing/same-agents-different-scores/index.md): A question from my Data Fest 2026 announcement: what still separates competition solutions when everyone has an agent? Archive date: 2026-05-28.
- [one matrix, many questions.](https://vk0st.com/writing/one-matrix-many-questions/index.md): One linear recommender, a short formula, and the modeling choices hiding inside it. Archive date: 2026-05-16.
- [two bars, one tiny screen.](https://vk0st.com/writing/flipper-codex-monitor/index.md): A small Flipper Zero side quest: putting Codex usage limits on a dedicated display. Archive date: 2026-05-13.
- [a qwen on the home card.](https://vk0st.com/writing/qwen-on-the-home-card/index.md): Qwen3.5-35B-A3B on an RTX 3090, running through llama.cpp on Windows and connected to Claude Code. Archive date: 2026-02-26.
- [a small conda trace.](https://vk0st.com/writing/a-small-conda-trace/index.md): A missing library, a short GitHub comment, and a few people getting unstuck in 2020. Archive date: 2020-06-01.
- [github](https://github.com/vk0st): Vlad's public profile.
- [kaggle](https://www.kaggle.com/vk00st): Vlad's public profile.
- [linkedin](https://www.linkedin.com/in/vladislav-kostoglodov): Vlad's public profile.
- [telegram](https://t.me/vkost): Personal Telegram contact.
- [candidate generation — Telegram blog](https://t.me/Vladik_repostnul): Vlad's niche blog; separate from his personal Telegram contact.