Information retrieval
VODLens CS2
Learn from the replay.
Describe a CS2 situation. Find comparable professional plays, with video evidence attached.
Take a closer look
An actual application screenshot from the project. Open it to inspect the details. The live retrieval service is not running on this portfolio.
01 / The question
A good example beats generic advice.
02 / What I built
I built a retrieval pipeline over professional match telemetry, comparing spatial, state-vector, semantic, fused, and LLM-assisted approaches. Video is joined to the retrieved moment as evidence.
The film room links each situation to a POV clip. Coaching briefs cite retrieved case IDs, and the application checks those citations before displaying them.
03 / What happened
Nine learned state weights improved nDCG@5 by 0.120 over the manually weighted state retriever. A single learned retriever then matched the three-way fusion within the reported uncertainty.
04 / Where it stops
The benchmark covers Mirage only: 1,928,108 situations, 145 match/maps, 70 queries, and 5,167 judged pairs. Relevance comes from an LLM judge checked against a small human-labeled set. The separate future-consistency metric agrees with the learned-weight gain’s direction, but does not clearly separate it. The reranker’s judged gain was not corroborated by that second metric.
05 / Keep exploring
One more? / 02