Skip to main content
AmigoFish

AI Engineer

AI EngineerVideoLast show:

Not rated yet

About this show

We turn high signal in-person events for the top AI engineers, founders, leaders, and researchers in the world into the best free learning opportunities for millions around the world here on YouTube. Your subscribes, likes, comments, speaking, attendance, or sponsorships goes a long way toward making our biz model sustainable indefinitely.

We strongly believe this industry deserves a better class of community and that we know how to do this well; we just need your support.

Videos come from the channel's public YouTube feed. The artwork is the newest video thumbnail. Last updated . Names, artwork and text belong to their publishers.

Cosmos

Similar by topic and creator

See the full Cosmos

Videos

  • AI Writes Code Faster Than Uber Can Merge It - Preetam Dwivedi & Manjari Akella, Uber

    · 27 min

    Preetam Dwivedi and Manjari Akella of Uber's developer platform team share how AI-driven code changes pushed Uber's merge pipeline to its limits, across monorepos where more than 5,000 developers land tens of thousands of commits a week. They walk through three changes to SubmitQueue that kept changes landing without long waits: admission gating that sorts cheap changes from expensive ones, a deferral queue that softens the impact on developers, and conflict relaxation that lets more changes run in parallel. They also explain how Stovepipe checks changes after they land and tracks the last verified-good state of trunk, why Uber moved from an always-green trunk to a hybrid approach, and what they're building next. Preetam Dwivedi https://www.linkedin.com/in/behinddwalls/ Manjari Akella https://www.uber.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!

    Not rated yet
  • Engineering Agency out of the Happy Path — Matthew Jewkes, Standard Cybernetics

    · 19 min

    A biopharma automation reached 99.95% accuracy on its business critical tasks after the team separated actual errors from ambiguous cases and exceptions. Matthew Jewkes describes building that system at Clarivate, where indexing drug development documents meant handling tens of thousands of new tasks a month. An early agent approach reached around 80% to 83%. Working through 1,000 evaluation examples with domain experts exposed three different problems: the model could be wrong, the organization could lack an agreed answer, or genuinely new science could require judgment. Each needed its own route. The routine path became a tightly specified automation, with uncertain and unusual cases escalated instead of pushed through. The larger payoff was recovering expert attention. Once routine decisions were handled, the remaining cases could be grouped for deliberate review: critical exceptions to clear, ambiguous low value work to freeze, and unusual demand that might justify a new product. One expert used that structured backlog to form an investment thesis before a major customer raised the same need. Jewkes lays out where the investment makes sense, how much evaluation work high reliability requires and why organizational ownership and trust matter alongside engineering. His bet is that companies will gain more from managing collective attention and judgment than from choosing a particular orchestration framework. Speaker info: - https://x.com/mjewkes - https://linkedin.com/in/mjewkes Timestamps: 0:00 - Expert attention and cheap execution 1:48 - Fully automate the happy path 4:18 - The biopharma indexing case study 6:39 - Errors, ambiguity and exceptions 8:52 - Routing uncertainty and the critical path 9:58 - From 95% to 99.95% accuracy 11:24 - Recovering expert judgment 12:51 - Exceptions as market signals 14:51 - Where full automation is worth the cost 17:17 - Evals and the value of human attention

    Not rated yet
  • I Let Agents Refactor My Codebase for 3 Weeks. Then I Read the Code. — Keiji Kanazawa, Microsoft

    · 19 min

    A coding agent wired unit tests to production inference, then ran them about 30 times. Keiji Kanazawa had spent three weeks letting coding agents refactor a personal research assistant. A bad digest sent him looking for a trace, and he discovered the "agent" was still a deterministic workflow. Skills had no SKILL.md, different architectural concepts were aliases for the same thing, and a daily reflection job was producing empty reports. He had approved plans without checking whether the implementation meant what its labels promised. He rebuilt around an explicit agent runtime, clearer specifications and goal rubrics, comparing four frameworks before selecting tools for his own needs as a product manager. The working example is a video deep dive agent that extracts 440 keyframes from an inference lecture, pairs blackboard equations with explanations and lets him question individual sections. The result generalizes to slide decks, coding demos and terminal recordings. He iterates on an evaluation rubric, lets agents optimize against it and uses adversarial reviews between coding agents. The rubrics need scrutiny too: one was tuned to a single lecture and failed on five other videos. His answer to the code reading debate is to own the goals and evaluation loop while continuing to delegate most implementation. Speaker info: - https://x.com/gojira - https://linkedin.com/in/keijikanazawa - https://gojira.blog/ Timestamps: 0:00 - Introduction 0:33 - The L/Z debate over reading code 1:56 - Building a personal research agent 4:00 - The agent that was not an agent 6:33 - Unit tests on production inference 7:59 - Rebuilding with clearer specifications 9:21 - Comparing four agent frameworks 12:04 - Video deep dives from keyframes 13:59 - Rubrics and evaluation loops 16:36 - Catching an overfit rubric 18:25 - Own the goals and delegate the code

    Not rated yet
  • What the Top Companies Do Differently With AI Coding - Martin Harrysson & Prakhar Dixit, McKinsey

    · 20 min

    Martin Harrysson and Prakhar Dixit of McKinsey, with their colleague Shilpa, share what they're seeing across large companies using AI agents in software development. Most still report modest gains, but roughly the top 20 to 25% are now seeing 2x end-to-end improvements. They walk through what those companies do differently: shorter agent-driven cycles instead of two-week sprints, smaller pods with merged roles, curated skills and shared context, and change management that goes beyond pilots. They also cover why quality can drop as code generation speeds up, how the bottleneck moves to requirements and release, and why measuring impact is still the hardest part. Chapters: 0:00 Intro 1:23 Coming out of the trough of disillusionment 3:32 Four patterns that work 5:10 Operating model: shorter cycles 7:15 Talent: smaller pods, merged roles 8:13 Tech: context and trust 9:24 Change management at scale 10:24 Curating skills 11:35 Faster code, lower quality 14:07 The bottleneck moves to requirements and release 16:07 Measuring impact 19:02 Wrap-up Martin Harrysson Prakhar Dixit https://www.mckinsey.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!

    Not rated yet
  • Not rated yet
  • How Ramp Tests Its AI Bookkeeping Agent Before Customers Do - Ryan Stevens, Ramp

    · 18 min

    Ryan Stevens leads applied science at Ramp and explains how the team built end-to-end evals for Stack, its AI agent for bookkeepers. He starts by breaking an accountant's job into tasks like reconciliation and variance analysis, then builds synthetic businesses with full books so agents can be back-tested on the same data again and again. He covers accuracy versus pass@k, why the latest models still fail mostly on reasoning, and the experiments that didn't work, including auto-tuning prompts. Then he shows the ones that did: removing unneeded skills, testing memories across two time periods, and scheduled regression runs that report problems in Slack. Chapters: 0:00 Intro 0:47 What's your agent's goal? 1:16 Breaking down a bookkeeper's job 2:43 Building a benchmark 3:20 Before and after you have users 4:26 Data is the bottleneck 4:48 Building synthetic businesses 5:45 Evals are just back-tests 7:08 How each task is graded 9:14 Accuracy vs. pass@k 10:24 Picking a model 11:18 Why models fail: reasoning 12:18 Experiments that failed 13:36 Skill ablation 15:01 Testing memories 17:01 Regression testing in the background 18:07 Takeaways Ryan Stevens https://ramp.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!

    Not rated yet
  • Claude Code Looked Right but Missed $800K - Natalia Connolly & Kevin Madura, AlixPartners

    · 18 min

    Natalia Connolly and Kevin Madura of AlixPartners pointed Claude Code at a real post-merger procurement analysis: hundreds of vendors and more than 40,000 contracts. The first answers looked complete, but one total was off by about $800,000 and a key renewal deadline was a month late. They walk through why: long contracts that silently failed to process, amendments that weren't linked to their parent contracts, and rate tables split across pages. They explain how a five-stage pipeline that structures the data first, plus custom tools with built-in business rules, made every answer traceable to its source. They also share lessons on clear access errors, de-duplicating savings found by parallel agents, and giving the agent context about what the business is already doing. Chapters: 0:00 Intro 0:52 The deal: two companies, 40,000 contracts 2:30 Answers that looked right 3:29 Why contract data is messy 5:46 A five-stage processing pipeline 7:08 Charts, tables and structured output 8:26 Custom tools with built-in business rules 10:20 Packaging what works 11:14 Lesson: the wrong date 11:59 Lesson: orphaned amendments 13:25 Lesson: access errors look like zeros 14:37 Lesson: de-duplicating savings 16:04 Lesson: know what the business is already doing 17:37 Takeaways Natalia Connolly Kevin Madura https://www.alixpartners.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!

    Not rated yet
  • How Financial Planners With Zero Coding Experience Built Their Own Agents - Matt Ragone, Range

    · 16 min

    Matt Ragone is a financial planner at Range, an all-in-one wealth management firm, not an engineer. He leads AI adoption for its planning team and shares the step-by-step method non-technical planners used to build their own agents and skills. With standardized processes and client data already in one place, the team could have an agent draft cash flow analyses, insurance reviews and tax reviews while planners are on calls with other clients, with a certified financial planner reviewing everything before it goes out. He walks through the five steps: pick a workflow you repeat, teach it real cases, package it as a skill, test it on edge cases, and ship it with an owner and a reviewer. Chapters: 0:00 Intro 0:42 A financial planner, not an engineer 2:28 Using AI as a chatbot 3:03 The bottleneck was typing speed 3:37 What changed with agents 4:21 What it took beyond a good prompt 6:27 Five steps to build a skill 6:58 Step 1: Pick a workflow you repeat 7:47 Step 2: Show it real cases 9:11 Step 3: Package it as a skill 10:50 Step 4: Test it on real cases 11:52 Step 5: Ship it with an owner and a reviewer 12:27 Where Range is now 13:29 Let the experts build the skills 14:28 The bottleneck has moved Matt Ragone https://www.range.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!

    Not rated yet
  • Your Coding Agent Can't Undo Its Own Mistakes - Tanmay Sah & Dolly Sah

    · 16 min

    Tanmay Sah and Dolly Sah look at a capability today's coding agents are missing: undoing what they've done. Drawing on their research projects UndoBench and EvoUndo, they show that an agent that completes a task can't necessarily put things back when something breaks, because Git doesn't cover external API calls, runtime config or notifications, and retrying doesn't reverse side effects. In EvoUndo, 197 of 478 changes that made an agent more capable (about 42%) also broke its ability to recover. They lay out an approach that captures the original state, picks the right response (retry, restore, compensate or escalate) and independently verifies the result. Tanmay Sah https://x.com/TanmayGradient Dolly Sah Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!

    Not rated yet
  • Your Company Has No Model of Itself. Agents Can Build One - Joao Wenzel, Citadel Securities

    · 18 min

    Every company has someone who knows everything: the reorgs, the politics, which projects are really on track. Joao Wenzel of Citadel Securities explains how his team approximates that person with an agentic system that builds an institutional memory from decks, Slack, meeting notes, code changes and agent traces. Borrowing the idea of a state estimator from robotics, the system updates its beliefs as evidence arrives, lets facts age, grades its own predictions against what actually happened, and takes corrections from people. He shows why simple retrieval isn't enough, how agents catch "watermelon projects" that are green on the outside and red on the inside, and how back-testing showed that projects the agent flagged were twice as likely to slip. Chapters: 0:00 Intro 0:22 The person who knows everything 1:16 Can agents help with strategy? 1:51 An institutional memory layer 2:18 Top-down and bottom-up evidence 3:05 Borrowing the state estimator from robotics 4:06 Getting data out of a status deck 6:11 The loop: update, age, observe, grade, correct 6:57 Why retrieval isn't enough 7:21 Update: building a record over time 9:12 Age: every fact has a half-life 10:12 Observe: specialist agents beat one big pass 11:12 Grade: back-testing what slips next 12:12 Correct: asking people the right questions 13:17 Bottom-up signals: Slack, PRs and agent traces 15:45 How Citadel Securities uses it 16:38 Open problems 17:24 Takeaways Joao Wenzel https://www.citadelsecurities.com Join us in SF for the second annual AIE CODE summit: https://aie.codes our last flagship summit of 2026. Applications live now!

    Not rated yet