Back office ops · Production

Deedy Das (Glean) on enterprise search architecture, the employee portal evolution, and why simple LLM-drop-in fails

The problem

Employees at most companies cannot efficiently find internal documents and knowledge the way Google employees can with internal tools like MoMA — knowledge is scattered across 10–100 SaaS apps with no unified retrieval layer, a gap that worsened with remote work.

First attempt

Earlier enterprise search attempts were on-prem and lacked modern SaaS API integrations, and more recent startups that simply drop in LLMs and embeddings still fail to produce quality results without rigorous ranking and tuning work.

Workflow diagram · grounded in source
1
SaaS data indexing via APIs
integration
“you also now have API support that's really nuanced on all of the SaaS apps that you use... a lot of SaaS apps have really robust APIs that really let index everything that you'd want”
2
Hybrid search query processing
ai_action
“we do a hybrid approach both using, you know, core IR signal synonymy. Query accentuation with things like acronym expansion, as well as stuff like vector search, which is also useful”
3
Personalization and ranking layer
ai_action
“our understanding of all of your interactions with people around you. Our personalization layer, our good work on ranking is what makes us good. It's not sort of, Hey, drop in LLM and embeddings and we become amazing at search”
4
Unified feed and mentions delivery
output
“all of the hundred Slack pings that you have, plus the Jira pings, plus the, the, the email, all of that in one place is super useful to have”
Reported outcome

Glean became a unicorn with customers including Databricks, Canva, Confluent, Duolingo, and Samsara, with users reporting they cannot go back to a working life without internal search.

Reported metrics
Stack Overflow traffic decline (attributed to ChatGPT, not Glean)15%
Reported stack
elasticLLMsSlackJiraGitHub
Source
https://www.latent.space/p/deedy-das
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Glean became a unicorn with customers including Databricks, Canva, Confluent, Duolingo, and Samsara, with users reporting they cannot go back to a working life without internal search.

What tools did this team use?

elastic, LLMs, Slack, Jira, GitHub.

What results were reported?

Stack Overflow traffic decline (attributed to ChatGPT, not Glean): 15% (source-reported, not independently verified).

What failed first in this deployment?

Earlier enterprise search attempts were on-prem and lacked modern SaaS API integrations, and more recent startups that simply drop in LLMs and embeddings still fail to produce quality results without rigorous ranking…

How is this back office ops AI workflow structured?

SaaS data indexing via APIs → Hybrid search query processing → Personalization and ranking layer → Unified feed and mentions delivery.

Related back office ops cases