{
  "video": {
    "id": "-uBABM2AuFg",
    "title": "Don't Approve the AI Model. Approve the App - Doneyli De Jesus, Langfuse",
    "duration": 1160,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": ""
  },
  "analysis": {
    "video_id": "-uBABM2AuFg",
    "title": "Don't Approve the AI Model. Approve the App - Doneyli De Jesus, Langfuse",
    "one_liner": "Regulated enterprises don't certify models, they certify applications, so build your own benchmark from a golden data set and gate every agent configuration change on per-dimension thresholds set by domain experts.",
    "summary": "Drawing on Langfuse's work with a top five bank, Doneyli De Jesus argues that public benchmarks don't translate to regulated firms' internal data and use cases, such as a 10-K and earnings call analyst for investment bankers. The process he shows runs a golden data set through the agent, keeps the resulting traces as evidence, and scores them across multiple dimensions to produce a pass/fail gate for production. In one example, a model scored 97% on groundedness but about 60% on numerical accuracy, and adding a deterministic calculation tool raised accuracy to over 90%, clearing the bank's 85% gate. Domain experts set the thresholds and review failures through annotation queues, and the bank's evaluations went from weeks to minutes.",
    "key_points": [
      "The bank told Langfuse directly that it certifies applications, not models, and public benchmarks may not translate to internal data and use cases.",
      "Finance deployments have three main requirements: on-premises (some banks are repurposing data centers and standing up GPU infrastructure to run models locally), gated from the internet, and air-gapped. Banks want to avoid even the appearance of a provider training on their data, despite zero data retention guarantees.",
      "The core mechanics are a golden data set of test cases (inputs, expected outputs and metadata), traces as evidence of what the agent was evaluated on, and scores that give pass/fail across selected dimensions. Banks often already have the golden data set but lack a way to operationalize testing against it.",
      "One model scored 97% on groundedness but about 60% on numerical accuracy. Giving it a Python calculation tool raised numerical accuracy to over 90% (about 95%), clearing the bank's 85% gate for that use case.",
      "Scoring blends deterministic code-based evaluators, path evaluation of whether the agent took the right route and skipped no steps, and LLM judges for qualitative aspects.",
      "Averages can hide fatal failures: a configuration can pass on groundedness and completeness yet score zero on regulatory compliance, a binary zero-or-100 dimension that can shut the gate on its own.",
      "Any change to the system prompt, model, tools or other configuration triggers a rerun of the gate, and results can be linked back to the specific change. End-to-end runs used data sets of about 150 and about 1,500 questions.",
      "Failures such as regulatory compliance misses are curated into annotation queues for domain experts, whose feedback improves the agent and supplies candidates for the golden data set. This helped the bank go from weeks to minutes on evaluation, in an area where risk and compliance review can sometimes take months."
    ],
    "takeaways": [
      "Write your own benchmark on your own data, starting with a golden data set treated like QA test cases, rather than relying on public benchmarks.",
      "Gate the application rather than the model, and rerun the gate on every prompt, model or tool change, so that swapping a model later is easier to adjust for.",
      "Set separate thresholds per dimension and per use case rather than passing on an average, and treat dimensions like regulatory compliance as binary blockers.",
      "Have domain experts, such as risk people on a risk use case, set the bar, then route failing traces to them through annotation queues and feed reviewed cases back into the golden data set.",
      "Give the model deterministic tools for tasks it is weak at, such as math, and combine deterministic, path and LLM-judge evaluators instead of relying only on LLM judges."
    ],
    "topics": [
      "evals",
      "agents",
      "observability",
      "regulated industries",
      "financial services",
      "quality gates",
      "golden datasets",
      "human in the loop"
    ],
    "tools": [
      "Langfuse",
      "ClickHouse",
      "Opus 5",
      "Python"
    ],
    "quotes": [
      {
        "text": "we don't certify models here. We certify applications, right?",
        "at": "03:08",
        "url": "https://www.youtube.com/watch?v=-uBABM2AuFg&t=188s"
      },
      {
        "text": "LLMs inherently are not great at math. They're just next token predictors.",
        "at": "08:03",
        "url": "https://www.youtube.com/watch?v=-uBABM2AuFg&t=483s"
      },
      {
        "text": "relying 100% just on LLM judges is not going to cut it.",
        "at": "12:09",
        "url": "https://www.youtube.com/watch?v=-uBABM2AuFg&t=729s"
      },
      {
        "text": "it might be that one dimension actually completely shuts down the gate for you to move this into production.",
        "at": "15:37",
        "url": "https://www.youtube.com/watch?v=-uBABM2AuFg&t=937s"
      }
    ],
    "words": 3836
  },
  "summary_url": "/#-uBABM2AuFg",
  "transcript": {
    "html": "/transcripts/-uBABM2AuFg.html",
    "txt": "/transcripts/-uBABM2AuFg.txt",
    "vtt": "/transcripts/-uBABM2AuFg.vtt"
  }
}