|
Forwarded this email? Subscribe here.
Presented by No Code MBA
👋 Welcome to The AI Builder Roundup
Voice agents need to hear, see, call tools, and still keep talking. Gemini’s new API models target that mix. Also: Perplexity’s faster search-storage pattern, typed AI decisions from Jev, and Salesforce data inside Claude.
In today’s edition:
- Gemini adds reasoning to live voice agents.
- Perplexity shares a faster search-storage design.
- Jev turns messy inputs into typed decisions.
- Salesforce brings CRM data into Claude.
🗣️ Gemini adds two lanes for live voice agents
Google has put Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking into public preview through AI Studio and the Gemini API. One is aimed at scalable, cost-efficient dialogue; the other targets harder multistep work without ending the conversation. Read the original post on X →
- Both models support near-real-time visual understanding, automatic language detection across 97 languages, and background tool calling.
- Gemini 3.8 Live is built for fluid dialogue, including mid-sentence interruptions and switching between languages.
- Live Extended Thinking can reason and speak in parallel, narrating progress as it works through a task.
- Developers get both models in public preview, while enterprise availability is private preview.
What it means: Build voice agents that can see context and call tools mid-conversation, then test the standard or Extended Thinking model against your latency and reasoning needs.
|
🗄️ Perplexity’s search store cuts batch-read latency
Perplexity says its in-house CobbleDB replaced DynamoDB for serving web content in search, dropping median batch-read latency from 31.4 ms to 5.60 ms and p99 from 123 ms to 24.2 ms. Its internal cost model estimates at least 20% savings relative to DynamoDB. Read the original post on X →
- CobbleDB is built for repeated batch reads of prepared page records, rather than as a general-purpose database.
- It groups page keys by partition and reads them in parallel with RocksDB MultiGet.
- Its router favors same-zone replicas and queries another copy when one replica is slow, preventing a delayed response from holding up the batch.
- Perplexity separated durable document state, update delivery, and serving so replicas can ingest updates independently and recovering nodes do not delay others.
- The company says it plans to open-source CobbleDB for teams building AI search.
What it means: For retrieval-heavy products, borrow the pattern: parallelize partitioned reads, hedge slow replicas, and decouple serving from update ingestion before adding more compute for latency.
|
⚡ Jev makes AI decisions look more like code
TypeSafe has opened early access to Jev, a model that takes unstructured input and returns typed, probabilistic decisions instead of generated text. The company is targeting scoring, routing, extraction, and branching jobs where an application needs a decision more than a chat response. Read the original post on X →
- TypeSafe describes Jev as a parallel model for structured outputs, with confidence and probability scores attached to each result.
- The company says calls take 70 ms to 500 ms, versus 3 to 329 seconds for the frontier-model comparisons in its evaluation.
- Its published workflow evaluation reports 193.6x faster and 444.6x cheaper results, but TypeSafe built the workflows and its comparison needs independent testing.
- Jev is in early access, with developers being brought in from a waitlist.
What it means: Put Jev on your shortlist for high-volume routing, scoring, and guardrail checks, but verify its calibration, latency, and economics against your own production data.
|
🤝 Salesforce data enters Claude’s workspace
Salesforce in Claude is now in beta, bringing accounts, opportunities, pipeline data, and 37 pre-built sales skills into Claude. Approved organizations can use it in Claude chat and Claude Cowork on web and desktop. Read the original post on X →
- The skills cover renewal preparation, QBR decks, pipeline coverage, and meeting follow-up.
- The beta is available on paid Claude plans for organizations Salesforce approves through its sign-up process.
- Organizations need access to the latest Sales Cloud enterprise edition to be eligible.
- Each person signs in with their own Salesforce account, so Claude sees only the data their existing permissions allow.
- Claude asks for approval before writing a proposed change by default.
What it means: Teams approved for the beta can test permission-aware CRM work in Claude while keeping individual Salesforce permissions and proposed writes under user approval.
|
👋 Thanks for reading. If you liked this roundup, forward it to a friend!
Want to advertise in this newsletter? Learn more.
This newsletter was created with the assistance of artificial intelligence. While we strive for accuracy, please note that AI-generated content may occasionally contain errors or biases.
To unsubscribe from this newsletter, click here.
|