# Work — Vlad Kostoglodov

Recommendation systems, from the model to the machinery.

## SANSA / same weights, less waiting

sparse models / performance.

Reuse residual columns and trim sparse computation while keeping the resulting weights unchanged.

My optimization of SANSA execution; draft PR in my fork.

2.04× faster fit · 17–18% less peak RAM · two datasets, 16 CPU threads.

Fit measurements on Books and Yambda-50M Like with 16 CPU threads and matched parameters. The work is in a draft PR.

[Read the technical account](https://vk0st.com/writing/sansa-less-waiting/index.md).

## Avito ML Cup / one matrix, several histories

retrieval / ranking / team solution.

Reuse learned relationships with different views of user history, then rank forty catalogue nodes.

My component was retrieval; Grigory Vyaznikov worked on ranking.

1st place · AmazMe · Recall@40 0.22375.

Team solution: candidate generation, history representations, and ranking forty catalogue nodes.

[Read the technical account](https://vk0st.com/writing/avito-ml-cup/index.md).

## VK LSVD / new video, familiar author

cold start / constrained ranking / team solution.

Find an audience through author history, then allocate users across videos under a shared budget.

Team solution; presented with Grigory Vyaznikov.

3rd place · Elves do RePlay · nDCG@100 0.37836.

Team solution: cold-start retrieval through authors and a shared user budget. The code archive includes experimental variants.

[Read the technical account](https://vk0st.com/writing/competition-rabbit-holes/index.md).

## RecSys 2026 / keep the evidence

dialogue recommendation / source attention / team solution.

Keep retrieval evidence separate through two candidate compressors, then explain the fixed ranking.

Team solution under AmazMe; the official entry is listed as vkost.

7th of 20 industry entries · Blind-B composite 0.5098.

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.

[Read the technical account](https://vk0st.com/writing/keeping-the-evidence/index.md).

## Riiid / five models, one prediction loop

sequential models / stateful inference / team solution.

Keep learner histories causal while combining two boosting models and three neural checkpoints.

My published inference notebook, with Dean as editor; transformer experiments retain teammate credit.

37th of 3,406 teams · silver · private AUC 0.806.

Published team inference notebook: Vlad is Owner, Dean is Editor. Transformer experiments retain teammate attribution.

[Read the technical account](https://vk0st.com/writing/riiid-transformer/index.md).

[Competition archive](https://vk0st.com/competitions/). [Talks and recordings](https://vk0st.com/talks/). [Projects and experiments](https://vk0st.com/lab/).

[Profile and technical work JSON](https://vk0st.com/facts.json).
[Human-readable selection](https://vk0st.com/work/).

Updated: 2026-10-14.