{
  "video": {
    "id": "l0FLhNqBOic",
    "title": "AI tools for Forward Deployed Engineering — Vasuman Moza, Varick Agents",
    "duration": 1223,
    "upload_date": null,
    "channel": "AI Engineer",
    "source": "AI Engineer",
    "event_format": "In person"
  },
  "analysis": {
    "video_id": "l0FLhNqBOic",
    "title": "AI tools for Forward Deployed Engineering — Vasuman Moza, Varick Agents",
    "one_liner": "Varick Agents' CEO and head of engineering argue that model execution is no longer the constraint — the bottleneck is understanding and re-engineering a customer's actual processes — and show the internal \"FDE agent\" they built (dependency graph + post-trained open-source model + RL-trained graph-traversal tools) to scale forward-deployed engineers without hiring 50 people.",
    "summary": "The talk's thesis is that with 2026-era models and harnesses, execution of knowledge work is close to solved, so the next bottleneck is going deep into a business, mapping how work actually happens (especially the failure paths), and redesigning it around AI. The CEO frames the forward-deployed engineer as the role that does this, and argues enterprises must get agents built on top of their systems of record — NetSuite, SAP, Salesforce, Dynamics — because they will never migrate off them. Head of engineering JD Pruitt then shows the internal \"FDE agent,\" a three-stage system (engagement agent, workflow agent, and a not-yet-built autonomous agent) built on a dependency graph of the company, a post-trained open-source model that writes tighter analysis than frontier models, and an RL environment that trains custom tools for traversing that graph to pull the right context.",
    "key_points": [
      "Execution is no longer the constraint: \"the models are improving to the point where intelligence is no longer the constraint,\" and knowledge work is \"almost entirely solved\" — what's left is designing how work gets completed around AI.",
      "FDE work has three parts: map how humans do the work today (critically, what happens when things go wrong — \"Sarah in AP handles the workflow today... but when things go wrong, she actually sends it over to Chris, who then takes four days of cycle time\"), re-engineer the process around AI, and deploy agents on top of existing systems.",
      "Re-engineering must be calibrated: different enough to capture ROI, but not so different that adoption craters — the worked example is an 8-step workflow where 4 steps go fully autonomous, 3 get human-in-the-loop, and 1 stays entirely human because the risk is too high or the value too low.",
      "Cited failure stats for the industry: an MIT review figure that 95% of generative AI pilots fail to reach production, and a second figure of 87% that don't produce measurable ROI — attributed to \"slapping AI onto broken processes.\"",
      "Enterprises won't leave their systems of record — one client spent $5,000,000 and 5 years migrating to NetSuite — so Varick OS spins up, monitors and governs agents (with an eval suite) that sit on top of Salesforce/NetSuite/Dynamics/SAP rather than replacing them.",
      "Point solutions (e.g. just prospecting, or just AP) deliver 5–10% ROI; department-wide transformation is claimed to deliver 25%, 50%, 75% in revenue uplift, cost savings and risk mitigation.",
      "The FDE agent has three stages: (1) an engagement agent — \"a better version of Claude built just for our FDEs\" that pulls in Granola notes, synthesizes documentation, reads PowerPoint slides and answers identity questions; (2) a workflow agent embedded in the platform that flags missed edge cases and process ownership while the FDE builds; (3) an autonomous agent, still in development, that would process client change-request emails and ship workflow changes without the FDE.",
      "Technical architecture: a dependency graph as the single source of truth (\"most of these workflows inside of enterprise are remarkably linear. They just have a lot of cycles in them\"), a model post-trained on open-source models (they're \"a fan of Kimiko 26\") to fix frontier models' verbosity, and an RL environment training custom graph-traversal tools — entity resolution (\"there's a lot of Mikes in every company we work with and Claude gets very confused by this\") and detecting redundancy cycles / DAG violations.",
      "The before state: FDEs were uploading ~150 pages of documentation to Claude, waiting ~2 minutes, and getting analysis that was \"verbose and incorrect.\""
    ],
    "takeaways": [
      "Interview process leads about the exception paths, not the golden path — documentation covers the happy path and one or two edge cases, but the handoffs and multi-day reconciliation loops are where AI actually breaks, and they differ company to company.",
      "Scope automation step by step across a whole department rather than shipping a point solution: assign each step to autonomous, human-in-the-loop, or human-only, and change the workflow enough to capture ROI without destroying adoption.",
      "Build agents on top of the customer's system of record; a pitch that requires migrating off NetSuite/SAP/Salesforce is dead on arrival in the enterprise.",
      "Don't accept frontier-model output for long client-facing analysis — post-train an open-source model for the detail/clarity balance consultants have, and treat context retrieval as a separate trained problem (RL over custom graph-traversal tools) rather than assuming a big knowledge graph is searchable.",
      "Before buying a graph DB, check whether Postgres and a plain dependency graph representation are enough."
    ],
    "topics": [
      "forward-deployed-engineering",
      "agents",
      "enterprise-ai",
      "knowledge-graphs",
      "post-training",
      "reinforcement-learning",
      "process-automation",
      "systems-of-record"
    ],
    "tools": [
      "Varick OS",
      "Claude",
      "Claude Code",
      "Codex",
      "Cursor",
      "Factory",
      "Anthropic",
      "Granola",
      "PowerPoint",
      "Slack",
      "NetSuite",
      "Microsoft Dynamics",
      "SAP",
      "Salesforce",
      "Postgres",
      "Kimiko 26",
      "MCP"
    ],
    "quotes": [
      {
        "text": "we you know upload about 150 pages of documentation to Claude and then we prompt Claude and then we wait like 2 minutes and then we get analysis and then it's verbose and incorrect and it kind of sucks.",
        "at": "12:44",
        "url": "https://www.youtube.com/watch?v=l0FLhNqBOic&t=764s"
      },
      {
        "text": "I will go as so far as to say that knowledge work is almost entirely solved.",
        "at": "10:43",
        "url": "https://www.youtube.com/watch?v=l0FLhNqBOic&t=643s"
      },
      {
        "text": "I only started believing in consultants once we started hiring them at Verek. And the reason is they're so good at figuring out what is the part of the detail the client actually cares about, and what is the part that can get glossed over. And frontier models have absolutely no concept of this.",
        "at": "16:38",
        "url": "https://www.youtube.com/watch?v=l0FLhNqBOic&t=998s"
      },
      {
        "text": "there was, you know, five companies trying to sell you a graph DB, and you can just use Postgres, whatever it is. Yes, I'm looking at you.",
        "at": "15:34",
        "url": "https://www.youtube.com/watch?v=l0FLhNqBOic&t=934s"
      }
    ],
    "words": 4503
  },
  "summary_url": "/#l0FLhNqBOic",
  "transcript": {
    "html": "/transcripts/l0FLhNqBOic.html",
    "txt": "/transcripts/l0FLhNqBOic.txt",
    "vtt": "/transcripts/l0FLhNqBOic.vtt"
  }
}