{
  "video": {
    "id": "1OMHGsUZiqA",
    "title": "How Forward Deployed Engineering is done at Kepler — Vinoo Ganesh",
    "duration": 1340,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "1OMHGsUZiqA",
    "title": "How Forward Deployed Engineering is done at Kepler — Vinoo Ganesh",
    "one_liner": "Vinoo Ganesh, who built Palantir's 'Project Frontline' FDE rotation program, argues forward deployed engineering is a product strategy — not a go-to-market role — and walks through four field stories that produced Foundry's ontology.",
    "summary": "The talk's single claim is that FDE is an extension of the product function, not the sales function: an FDE is judged by their ability to identify areas of opportunity on site and generalize product solutions out of them. Ganesh tells four Palantir stories — a Cassandra keyspace blowup from building in isolation, a 47-page requirements doc that was really a Slack alert, a Parquet migration blocked by a missing file viewer, and a throwaway Groovy script that ran for years — each producing a rule for what to detect, where to be, what to name, and what to ship. The ontology (nouns = entities, verbs = operations) is presented as an accident Palantir tripped into while trying to let humans work in their own domain language, and as the thing that later let go-to-market FDEs simply sell and integrate Foundry.",
    "key_points": [
      "Phoenix (2013) was built in isolation without talking to financial customers; blank date values defaulted to the epoch, and the 10-minute time-bucketed retention scheme between 1970 and 2013 generated 2.3 million keyspaces — at Cassandra's 5MB per file handle that meant 14TB of RAM just to start the server, dead on arrival.",
      "A shipping customer's VP of operations produced a 47-page requirements doc for a 14-metric dashboard with drill-downs and alerts, scoped as a three-month dev project; four months of scoping later, one engineer visited on site, asked what the dispatcher does Monday morning, and shipped a Slack alert in 4 hours.",
      "A data quality engineer fought a CSV→Parquet migration for a year, saying 'Parquet is way worse'; on-site observation showed she was downloading CSVs from S3 to a Windows box and double-clicking them to spot-check quality. They built a Parquet viewer that night, she approved the migration in 2 days, and pipeline execution went from 17 hours to about 2.",
      "Concrete on-site tells that mark an opportunity: any task a user does more than once, copy-paste between tools, tool or tab switching, the phrase 'well, I have to', and pulling out a phone mid-task. Opening question: 'what's the most annoying part of your morning?'",
      "Ontology came from terminology collisions — sales says customers, ops says clients, finance says billing entities, devs say org IDs — and the same overloading applies to DAU (high-quality users for product vs. anyone who logged in for infosec) and today to 'skills', 'MCPs' and 'agents'.",
      "A Groovy script hacked together as a temporary retention fix ended up running everywhere at a ~100,000-person customer, nicknamed 'vinoo.groovy' — people wore it on shirts at his wedding — and had to be supported for years because it was never productized.",
      "Ganesh built Project Frontline, the rotation that converted Palantir software engineers into FDEs; ~350 people went through it, and he built the same program at Citadel for portfolio-manager data products.",
      "Foundry was a general product solution informed by tens of thousands of hours in the field; becoming the linguistic foundation is what locks you in, because users adopt your language along with your product."
    ],
    "takeaways": [
      "If solving the problem takes under a day, just build it and ship it and close the loop — don't expand it into product vision or bring in PMs. Whoever defines the problem owns the solution.",
      "Before building, answer three questions: what are you trying to accomplish, what happens after you have the solution, and how is the customer solving it today. Customers describe solutions, not problems; the FDE's job is to find the problem and define what happens next.",
      "Get badged on site — 'residents get the truth'. Your customer-site badge and contractor email address are your data mining permits; you can't survey your way to the real workflow from a conference room in New York.",
      "Extract the ontology deliberately: which nouns are overloaded, where the integration points and translation layers are (Snowflake→Databricks, Palantir→Tableau, Anthropic→SAP), which systems of record can never be swapped, and what words people use when describing their issues.",
      "Calibrate what you ship with four questions: will I get a 2am call about this in 6 months, what's the trade-off of solving it fast, who do I hand it off to, and what happens when it breaks. Ship everything as if it will run for 18 months, because it probably will."
    ],
    "topics": [
      "forward-deployed-engineering",
      "product-strategy",
      "ontology",
      "enterprise-software",
      "customer-discovery",
      "data-engineering",
      "technical-debt"
    ],
    "tools": [
      "Palantir",
      "Foundry",
      "Phoenix",
      "Cassandra",
      "Parquet",
      "S3",
      "Slack",
      "Groovy",
      "Elasticsearch",
      "Snowflake",
      "Databricks",
      "Tableau",
      "SAP",
      "Kepler",
      "Citadel",
      "OpenAI",
      "Anthropic",
      "xAI"
    ],
    "quotes": [
      {
        "text": "the first fundamental truth of FTE is that this is not a role this is a product strategy",
        "at": "02:15",
        "url": "https://www.youtube.com/watch?v=1OMHGsUZiqA&t=135s"
      },
      {
        "text": "Your badge on site at a customer site and your email address, your contractor email address, those are your data mining permits.",
        "at": "11:55",
        "url": "https://www.youtube.com/watch?v=1OMHGsUZiqA&t=715s"
      },
      {
        "text": "when a user adopts your product, they also adopt your language. How many people are calling things skills right now? How many people are calling things MCPs that are function calls with prompts?",
        "at": "14:36",
        "url": "https://www.youtube.com/watch?v=1OMHGsUZiqA&t=876s"
      },
      {
        "text": "The most dangerous words in forward deployed engineering or engineering is this is just temporary.",
        "at": "19:14",
        "url": "https://www.youtube.com/watch?v=1OMHGsUZiqA&t=1154s"
      }
    ],
    "words": 4641
  },
  "summary_url": "/#1OMHGsUZiqA",
  "transcript": {
    "html": "/transcripts/1OMHGsUZiqA.html",
    "txt": "/transcripts/1OMHGsUZiqA.txt",
    "vtt": "/transcripts/1OMHGsUZiqA.vtt"
  }
}