Scope searches with security built in to solve document access problems
Vinoo Ganesh recommends that the way to actually solve these document access problems is scoping searches with security built in.
35,595 results · Most relevant
Vinoo Ganesh recommends that the way to actually solve these document access problems is scoping searches with security built in.
Vinoo Ganesh recommends creating a system where the models have to dereference their own pointers, saying that by doing so you can actually create a very powerful, very verifiable AI platform.
For any tool you use, ask whether you can actually verify the numbers it produces. Ask whether you can see where the sources are — not as after-the-fact annotation, but by clicking on the number, whether it's a calculation, an extraction, or a combination — to make sure you trust the insight.
Dziri advises using a strong lead agent with cheap worker agents to get near-frontier results at a fraction of the cost, noting the team does not have to be all strong models.
Splitting work among a team of agents lets each worker start with a fresh context, read files, run tests, dig through logs, and send back only a compact result. The lead keeps the full picture and can finish faster if the task is decomposable and agents can work on independent pieces.
Dziri advises the audience to find the bottleneck in their agent pipeline and scale the team there, arguing that where you put agents matters more than how many agents you have.
Takeaway 2: Where you put agents matters more than how many — find the bottleneck and scale the team there.
Takeaway 3: The team does not have to all be strong — a strong lead with cheap workers can get near-frontier results at a fraction of the cost.
Frame team design as a compute allocation problem: ask where the next dollar of inference buys the most reward over the solo agent.
Jesse Hall recommends that when a voice agent's backend tool call fails or errors, the agent should self-correct by verifying the information with the caller — saying something like 'let me verify this' — rather than silently retrying or stalling. When the agent cannot resolve an issue, it should escalate to a human as the worst-case fallback rather than continuing to fail.
Point every model call at a privately hosted LLM to keep sensitive data local. Design each model call as small, typed, and schema-checked so switching to a privately hosted model is a configuration change rather than a rewrite.
Rob Snyder recommends reframing selling in your head: you're not convincing anyone to buy — you're just seeing if they need help with anything they're stuck on. If they don't, that's fine; if they do, you can explore whether you can help. He admits he was nervous to sell as a mega introvert and overthinker, and this reframe made the question comfortable. The host adds that sometimes you can simply ask what problem you can solve for them.
None stated.
Rob Snyder identifies the main brutal anti-pattern in idea generation as feeling like you need to build it yourself first — going heads-down and building before talking to customers. He advises facing the cold wind of reality as soon as possible by not delaying customer conversations where pull can be found. He notes that with AI you can get the build-it-first urge out of your system in a couple of weeks instead of months.
The AI timeline claim is described with moderate confidence.
Rob Snyder recommends that once a couple of people say yes, don't just throw the product over, send them a login and pray — go deliver the product firsthand as well.
Rob Snyder recommends designing the onboarding process to the leading indicator of retention (X), around the expectation that customers should not churn.
Sofia Jakovcevic recommends selecting a dataset by pulling raw real production traces instead of synthesizing data, pulling real user queries, real inputs, and real tool calls directly into the dataset.
Sofia Jakovcevic recommends bringing in evaluators as a scoring mechanism for the regression suite, automating subjective agent metrics like hallucination, accuracy, tool trajectory, and step efficiency into numeric scores. She uses LLM-as-a-judge to judge a span, trace, or log data against an LLM with a guide for determining hallucination or efficiency, and advises using code-based evals for deterministic checks and LLM-as-a-judge for non-deterministic checks.
Sofia Jakovcevic recommends running a baseline experiment by running the agent against all production traces and then running the eval testing suite on the output. In the SDK, you pass in the dataset, task, and evaluation suite to get results. You can adjust the policy, rerun, and compare how changing the prompt, handoffs, tools, and models affects overall output, all without deploying to production.
The name of the game for quantization quality is rigorous evaluation loops: benchmarking, agentic evaluations, and a lot of in-house dogfooding. Run a measurement-quality red-teaming loop before sending any model to any user, internally dogfooding the model for quite a bit first. Internal dogfooding catches tool-calling regressions (if your cloud code harness can't search the internet, you'll notice it as a user) and catches when the model is not as smart as it should be, via benchmarks and end-to-end evaluation scores.
Write very small atomic rewrite rules that specify different equivalencies, then let the search figure out what to apply when. Do not manually tell the search which rewrite rules to apply.
Choose sources → to get a library of your own and a daily brief.