The setting
supplies cloud software that lets developers and businesses build messaging, voice, and email into their applications. Its gathers these tools, along with identity checks and customer data.
Current preferences need priority, but historical memories still have a purpose.
Filtering old preferences protects immediate accuracy, while ranking them preserves answers about past conversations.
Who you’d be doing this for
“Our tester changed her delivery time twice, and the agent still offered the morning slot she'd rejected.”
Syafiq Aung · Application Developer · a character in this brief
Maintains a retailer's conversational application using retrieved preferences to suggest delivery times.
What is at stake
Agents received outdated preferences first in 36 of 200 requests, despite correctly stored updates. You weigh a targeted ranking correction against a filter that excludes superseded preferences from retrieval.
Why it isn’t already fixed
Every obvious fix costs something else. That’s the part you’d have to decide.
- current accuracy vs historical access
- ranking sensitivity vs retrieval latency
- targeted feature correction vs narrower retrieval scope
Why Twilio
Conversation Memory retrieves customer context in milliseconds, so an accurate stored update is insufficient when the ranking puts its predecessor first.
Written with these in mind
Not your kind of problem? 41 more at Twilio, or browse every organization.
Where the organization context comes fromThe postings this brief was written from, and the reporting read before it.
The organization context comes from public information. The situation and characters are written for the brief.
Job postings it was written from
Public descriptions of the work at Twilio. They are listed as sources, not as openings. The lines under a posting are its own words, where it meets this brief.
- Staff Machine Learning Engineer (opens in a new tab)
- Machine Learning Engineer 2 (opens in a new tab)
- Own the retrieval path that answers "what do we know about this person?" in milliseconds, working with embeddings, vector search, and ranking
- Build the extraction pipeline that turns raw data into structured memory, tuning the LLM prompts and schemas that decide what an AI agent remembers about a customer and what it forgets.
This is the setup. The work is inside.
Running it puts you in the room: the full situation and its constraints, stakeholders who push back in their own words, and the decisions that are yours to make. What you produce becomes a Day One Plan — work you can show someone instead of describing.
Where similar work appears
Other live briefs connected to this one, and what connects them.
Similar problems at other organizations
Like this one, each of these turns on fixing one thing without breaking another.
P95 first response has risen from 1.1 to 2.4 seconds after redundant retrieval calls appeared.
You must weigh a faster cache path against session-safe context and answer quality.
Clinician review found outdated GLP-1 guidance in 9% of sampled follow-up messages.
You must weigh a fast retrieval correction against response speed and reliable regression coverage.
Clinicians are rejecting exercise-progression recommendations because the agent retrieves outdated protocol content.
You must weigh a fast retrieval correction against regression risk across shared deployments.
The same setting, from another role
Each is set around Twilio’s Twilio Customer Engagement Platform, written for a different role. They’re separate briefs, not the same scenario.
Sellers found conflicting buyer assumptions in eight of twelve Messaging deals marked ready for proposals.
You decide how a draft handles a country or volume figure when an Outreach email differs from the Salesforce record, and which differences go to the seller first.
Five software providers choose a supplier in eight weeks, but disagree on who should bill their customers.
You weigh partner incentives and customer ownership against the margin and support obligations each agreement creates.
A software partner’s $9 million Messaging renewal expires tomorrow amid disputed forecasts.
You must weigh retaining its full commitment against signing away the ability to recover rising carrier costs.
Other problems with a similar purpose
In each of these, as in this one, the aim is to make a machine-driven decision dependable and fair at the exact real-world moment a customer is counting on it.
Fourteen percent of Qwen requests now exceed three seconds after the container build.
You must weigh faster batching against the context available to each answer.
Duplicate medication events appear in 6.8% of amended-record cases and inflate treatment timelines.
You have to weigh fast suppression against losing real repeat prescriptions.
Thirty-one legitimate internal messages were blocked in one day after reporting lines changed.
You must weigh a fast feature correction against a broader exception path that consumes the same release window.
