Platform PM · Prezent.ai

Search and Discovery

Built a novel slide search and cut "no results" from 10% to under 1%.

The stakes

A $1.2M account was frustrated: their teams couldn't find slides they knew existed. On a content platform, an asset you can't find might as well not exist.

The job

Redefine what it means to search a slide, then make it work for 150+ enterprise accounts without slowing anything down.

How

  1. 01

    Define

    What makes a slide findable: imagery, text, visual flow, business concept.

  2. 02

    Extract

    Iterated from YOLO to BLIP to Gemini Flash for rich, reliable tags.

  3. 03

    Store

    Tags and embeddings written to a vector DB inside the upload pipeline, so search latency never takes the hit.

  4. 04

    Retrieve

    Apache SOLR runs hybrid retrieval: exact keyword match plus semantic embedding search.

  5. 05

    Rank

    Relevance tuned to the audience: does this user want visuals or data?

  6. 06

    Evaluate

    Benchmarked on 1,000 real customer queries: 99% solved, confirmed by product analytics.

<1%
"No results" rate, down from 10%
99%
Of 1,000 real queries solved in eval
$1.2M
Account saved