Engineering Alpha in Private Equity

Episode

• 
  Engineering Alpha in Private Equity Trailer 01.05.2026 3mnt

  Welcome to Engineering Alpha in Private Equity, the podcast about how software engineering and data science excellence create operational alpha in private equity. In this trailer, Dave and Paul discuss the founding thesis of the show and what listeners can expect. ## Topics covered - What "operational alpha" actually means - Who are Paul Karner and Dave Mangot - What you can expect from the show - The best ways to engage [Learn more](https://engineeringalpha.fm)
• 
  What's Engineering Alpha & Why You Can’t Just Rub AI on It 05.05.2026 28mnt

  Welcome to the first official episode of the Engineering Alpha and Private Equity podcast! Hosts Paul Karner (economist and data scientist) and Dave Mangot (DevOps expert) break down what "Engineering Alpha" actually means for middle-market PE firms. Moving past the old world of financial engineering, Paul and Dave explore how true operational excellence within engineering organizations drives outsized returns and high EBITDA. They also tackle the elephant in the room: AI. They explain why it's a tool, not a magic product, and why failing to build the right foundations will amplify your problems rather than your profits. Key Takeaways: - Defining Engineering Alpha: Why optimization and efficiency inside the engineering organization are the true drivers of operational leverage and higher ROI. - The Deming Philosophy: How W. Edwards Deming’s statement that 94% of problems are systemic (and thus management's responsibility) applies directly to PE investors and C-suite executives. - AI Reality Check: Why AI is an amplifier for both good and bad processes, and why you shouldn't just mandate an "AI story" from the board without establishing the foundations in the DORA research. - The CircleCI Report: Exploring recent data showing that while AI helps developers write more code, it's often trapped in feature branches, has longer outages, higher customer churn, and negative ROI. - The New Valuation Metric: Why "agentic proficiency is the new SaaS multiple" and how building scalable foundations improves unit economics and drives higher valuations. Links & Resources Mentioned: - [DORA](https://dora.dev/) (DevOps Research and Assessment) State of DevOps Reports and the Accelerate book - [CircleCI Report](https://circleci.com/resources/2026-state-of-software-delivery/) on continuous integration tests - Nick Lichtenberg's [Fortune interview](https://fortune.com/2026/04/28/tech-layoffs-ai-disruption-corporate-america-doesnt-one-silicon-valley-ceo-knows-why/) with the CEO of Box - [Agentic Proficiency - The New Premium SaaS Valuation](https://blog.mangoteque.com/blog/2026/04/15/agentic-proficiency-the-new-premium-private-equity-saas-valuation/) Podcast theme music by [J-KIND](https://soundcloud.com/jkind). Connect with [Paul](https://www.linkedin.com/in/pkarner/) and [Dave](https://www.linkedin.com/in/dmangot/) on LinkedIn to join the conversation. [Learn more](https://engineeringalpha.fm)
• 
  Finding “Free EBITDA” in Cloud Contracts & The AI Optionality Playbook 14.05.2026 17mnt

  In this news review episode, we break down the recent wave of partnerships between major cloud vendors and private equity firms, starting with the Thoma Bravo and Google Cloud announcement. These partnerships highlight an immediate lever for value creation: enterprise cloud agreements that can drastically reduce operating expenses and instantly boost P&L. Beyond the immediate cost savings, we explore the strategic necessity of maintaining “optionality” in a highly uncertain AI landscape. We also issue a warning to CTOs: stop isolating your solutions architects in “innovation labs” and expecting new technology to fix broken systemic problems. Key Takeaways: - The “Free EBITDA” Play: Why your portfolio companies are leaving money on the table if they aren’t negotiating enterprise agreements with AWS, GCP, or Azure. Dave shares a real-world example of securing a 50% discount on internal bandwidth costs. - Why AI Optionality is King: In a highly volatile AI market, getting locked into a single LLM vendor is a massive risk. We explain why the best operational playbook involves using cloud platforms to access multiple models (like Anthropic, DeepSeek, and Gemini) to build a custom “race car”. - The Solutions Architect Trap: Why bringing in solutions architects to build a segregated “skunkworks” or innovation lab is a recipe for failure. - Tech Can’t Fix a Broken Org Chart: If your development team and your SREs report to different executives with misaligned incentives, no amount of AI or cloud architecture will help you hit your exit targets.
• 
  Uber COO questions tokenmaxxing 26.05.2026 6mnt

  In this hot-take episode, Paul Karner and Dave Mangot analyze a massive red flag in the current tech landscape: Uber's COO recently admitted it's getting harder to justify the money spent on AI "token maxing," while the company is slowing hiring to fund these AI investments. Dave and Paul break down why high token consumption often just creates stranded "inventory" rather than revenue-generating features, and why cutting your engineering labor force before AI proves its actual ROI is a massive mistake. Key Takeaways: - The "Token Maxing" Illusion: Why an increase in AI token consumption does not proportionally translate into useful consumer features or revenue. - Inventory vs. Revenue: AI helps developers write more code, but it's getting stuck in feature branches. That code is simply "inventory," and you only make money when inventory hits production. - Protecting the Productivity Engine: Why the decision to slow headcount/labor to offset AI costs is deeply flawed if the AI isn't actually yielding the expected efficiency gains. - The Data-Driven Playbook: Why leadership must look at actual production metrics and token ROI before disrupting the underlying labor force. https://www.businessinsider.com/uber-coo-andrew-macdonald-ai-token-spending-harder-justify-2026-5
• 
  Measuring the ROI of Tech Transformations 28.05.2026 23mnt

  How do you actually prove that a massive technology transformation is working? In Episode 4, we flip the script and dive into Paul Karner’s world of data science and causal inference. Using a real-world example of a private equity portfolio company that implemented machine learning to automate manual document review, Paul explains the critical process of tracking tech ROI. Dave and Paul break down why a tech transformation shouldn't just be a cost-saving measure, the nightmare (and necessity) of tying operational tech data directly to your ERP, and how to start with your financial end-goals and work backward. Key Takeaways: - The Cost-Savings Trap: Why undertaking a tech transformation where the only goal is cost savings usually leaves money on the table—or fails entirely. - Connecting the Silos: Why you must marry up operational data with your timekeeping and ERP systems to prove that a new tool is actually saving labor hours. - True EBITDA Leverage: How to avoid the "tokens versus humans" trap. Instead of just replacing headcount, learn how redeploying saved labor hours to scale output sends every additional margin dollar straight to the bottom line. - The 30-Day Playbook: Paul’s actionable advice for leaders currently in a transformation: start with a clear vision of how the project rolls up into the P&L, crack into the financial data, and work backward to find the gaps.
• 
  The AI Amplifier Effect & Why Rework is Killing Your EBITDA 04.06.2026 50mnt

  In our first dedicated research episode, Dave and Paul are joined by Nathen Harvey, who leads the DORA (DevOps Research and Assessment) program at Google Cloud. For over a decade, DORA has proven that elite software delivery performance is a leading predictor of organizational performance, including profitability and market share. In this episode, we unpack DORA's latest findings on Artificial Intelligence to show operating partners exactly why "rubbing AI" on a portfolio company won't work without the right foundational capabilities. Nathen breaks down the dangerous reality of AI as an amplifier: if you speed up code generation but ignore your existing bottlenecks, you will actually reduce your overall output.
• 
  The 'Maintenance Window' Red Flag & The AI Death Spiral 11.06.2026 15mnt

  If a software company still uses "maintenance windows" to release updates, it is a glaring operational warning sign. In this explainer episode, Dave and Paul break down why maintenance windows indicate a broken software delivery culture that relies on subjective feelings rather than automated data. For private equity operating partners evaluating a new acquisition or monitoring a portfolio company, Dave explains why legacy practices like Change Advisory Boards (CABs) actually decrease stability. More importantly, the hosts reveal why trying to force AI tools into an organization that deploys slowly will create a margin-crushing "death spiral" of dual costs. Key Takeaways: - The Legacy Tech Tax: Why maintenance windows signal that a company lacks automated testing and relies on subjective measures rather than objective facts. - The AI Death Spiral: If you use AI to generate 10x more code, but only release during scheduled windows, you are paying for AI tokens and paying engineers to perform massive amounts of rework when those giant batches fail. - The CAB Illusion: Why Change Advisory Boards (CABs), often used for compliance in highly regulated industries, are actually inversely correlated with software stability. - Killing Product-Led Growth (PLG): You cannot execute a PLG strategy without running continuous, daily experiments to see what customers want. Maintenance windows actively choke off this growth engine.
• 
  Double-Speed Baseball and the new SaaS Multiple 18.06.2026 14mnt

  In Episode 7, Paul Karner and Dave Mangot unpack a core thesis of the show: why "agentic proficiency" is the modern equivalent of the SaaS multiple. While the ultimate goal of any portfolio company remains the same — satisfying user needs to generate revenue and EBITDA— the mechanics of how we deliver that value have fundamentally changed. Dave and Paul break down how AI agents drastically lower the marginal cost of delivering software and why achieving daily deployment is the absolute prerequisite for Product-Led Growth (PLG). The message is clear: if your portfolio companies aren't using agentic workflows to get more "at-bats" in the market, they are going to be left behind by the compounding advantages of elite performers. Key Takeaways: The New Valuation Lever: Just as the industry previously rewarded the shift from legacy architectures to the cloud with massive SaaS multiples, the next wave of outsized exit multiples will go to organizations that master agentic proficiency. Unblocking Product-Led Growth (PLG): You cannot execute a successful PLG strategy if your ability to ship software is slower than your ability to learn what the market wants. Agentic proficiency removes the shipping bottleneck, allowing product teams to iterate daily. Double the "At-Bats": If two portfolio companies are competing to generate EBITDA and revenue, the agentically proficient company gets twice as many opportunities to deploy revenue-generating features and capture market share. The Compounding Advantage: Getting 1% better every day through continuous, agent-assisted shipping creates a compounding effect. This operational leverage causes elite organizations to drastically diverge in valuation from competitors who are merely trying to bolt AI onto legacy systems. https://www.antmurphy.me/newsletter/fix-delivery-first
• 
  Token Economics, Data Moats, and the Future of Tech Due Diligence 23.06.2026 48mnt

  In Episode 8, Paul Karner and Dave Mangot are joined by Dan Bender, Kirby Montgomery, and Jason Langenauer from the global tech due diligence firm Code & Co. (https://www.codeandco.com/) The team breaks down how the rise of AI has fundamentally changed the diligence process for private equity investors. The conversation shifts away from the hype of AI and dives straight into the P&L consequences of "token economics". The Code & Co. team explains why blindly throwing AI at a problem will wreck a software company's gross margins, why proprietary data is the only genuine moat left in the age of commoditized coding, and why a CTO's cultural skepticism toward AI is now considered a material investment risk. Key Takeaways: The Death of Zero Marginal Cost: Why the traditional SaaS model (where adding a new user costs almost nothing) is dead if your portfolio company is burning through expensive LLM tokens for every transaction. Token Economics & P&L: How to prevent margin erosion by matching the right AI model to the right problem (e.g., using a fraction-of-the-cost model like Haiku for basic tasks instead of the most expensive models). The Data Moat: Software development is becoming commoditized; clean, proprietary data is the only true competitive advantage that competitors cannot replicate. The 12-Month Sell-Side Shift: Why operating partners need to shift their sell-side tech diligence left, looking at the plumbing 12 months before going to market to build a convincing narrative around AI defensibility.
• 
  The AI "Hoax," Economic Accounting, and Nobel-Winning ROI 29.06.2026 12mnt

  In Episode 9 Paul Karner and Dave Mangot tackle a recent Fortune interview (https://fortune.com/2026/06/21/nobel-laureate-daron-acemoglu-ai-productivity-capitalism-democracy/) with Nobel Prize-winning economist Daron Acemoglu, who argues that the massive productivity gains promised by AI are harder to achieve than presumed. Paul puts on his PhD economist hat to break down what this skepticism means for private equity deal teams trying to manage their AI budgets. He introduces the concept of "economic accounting"—understanding the counterfactual of what a company could achieve without AI simply by adopting solid engineering foundations. For operating partners, the takeaway is clear: preparing a portfolio company for AI requires cleaning up data, mapping workflows, and establishing guardrails. Even if the AI hype is overstated, doing this foundational work will inherently make the company more profitable. Key Takeaways: The AI "Hoax" Analogy: Preparing for AI forces organizations to implement best practices. Even if AI doesn't yield AGI-level miracles, those foundational improvements directly increase EBITDA and profitability. Economic vs. Financial Accounting: CFOs must look at the "counterfactual": evaluating what productivity gains are actually coming from the AI versus what gains are just the result of getting the company's operational house in order. Ending the Token Free-For-All: Moving from a subsidized "token-maxing" phase to a mature operating model requires pointing a sustainable budget only at areas where AI truly creates unique value. Multiple Expansion: Deal teams that stop blindly "rubbing AI on everything" and strategically direct dollars toward genuine tech efficiencies will see the results directly in their exit multiples.
• 
  The 85% Inventory Trap: What 28 Million Workflows Reveal About AI ROI 09.07.2026 29mnt

  In Episode 10 of Engineering Alpha in Private Equity, Paul Karner and Dave Mangot dive into the hard data from the 2026 CircleCI State of Software Delivery Report, which analyzed over 28 million CI workflows. While the tech world is celebrating a 59% increase in code throughput due to AI, Dave and Paul reveal a massive P&L red flag: 85% of that new code is getting stuck in "feature branches". This means the AI isn't generating operational alpha; it is generating expensive, unsold inventory. They break down why only the top 5% of elite engineering teams are actually pushing this code to production, why test failure rates are skyrocketing, and why companies are accidentally paying the equivalent of multiple full-time engineers just to debug AI errors. Key Takeaways: The Feature Branch Inventory Trap: Code stuck in a feature branch doesn't generate revenue. It is expended capital sitting as inventory. You only make money when that code ships to production. The 30% Failure Tax: Because AI generates code so quickly, test success rates have plummeted from 90% down to 70%. For a high-throughput portco, that equals an additional hundreds of hours of debugging every year—the equivalent of many full-time engineers doing nothing but fixing AI mistakes. Kill the Vanity Metrics: Boards must stop measuring "lines of code" or AI adoption rates. The bottleneck is no longer how fast developers can work; it is whether the underlying systems can keep up and safely deploy that work. The Elite 5% Divergence: Only the top 5% of software teams have the foundational systems required to actually capture the promised ROI of AI, successfully shipping 25% more code to production. https://circleci.com/resources/2026-state-of-software-delivery/
• 
  3 a.m., the 2x Productivity Cap, and the ceiling of AI Automation 28.07.2026 23mnt

  In Episode 11 of Engineering Alpha in Private Equity, Paul Karner and Dave Mangot react to a recent Anthropic ad showcasing Claude Code automatically fixing a production bug at 3 a.m. While the ad promises a "find it, fix it, and ship it" utopia, Dave and Paul deconstruct why this is a dangerous fantasy for most mid-market PE-backed companies. They reveal why CTOs who have trained their engineering teams on AI are hitting a hard "2x productivity cap" — generating code faster, but failing to actually ship it. This episode is a roadmap for the actual systemic changes (like platform engineering and automated testing) required to break through that cap and deliver the financial ROI promised to the board. Key Takeaways: The 2x Productivity Cap: Why simply handing your engineers AI tools will cap out at 2x productivity. Without systemic redesigns, engineers just write code faster, piling up expensive, unshipped "inventory". The Hidden Prerequisites: Agents cannot fix your systems if they can't read your logs. The Anthropic ad accidentally proves that elite platform engineering (like robust Kubernetes environments) is the mandatory foundation for AI success. The Submarine Rule (Is it Safe?): Why you should never let an AI agent independently "find it, fix it, and ship it" to production. Elite engineering leaders treat AI like a submarine crew: the agent must propose a fix and explain why it is safe before a human authorizes the deployment. Delivering Board Promises: The foundational building blocks discussed in this episode are the exact investments required to get past the 2x plateau and actually deliver the EBITDA gains promised in the investment thesis.
• 
  Doing Math on Opinions: Why Bug Counts Aren't Science 05.08.2026 45mnt

  Paul and Dave sit down with tech veteran and quality pioneer Elisabeth Hendrickson (former VP of R&D for Data Products at Pivotal). Together, they shatter one of the most common metrics used in private equity boardrooms: the bug count. Elisabeth explains why tracking and graphing bugs is actually just a statistical illusion—doing "math on human opinions" rather than measuring actual technical quality. The team discusses how quality directly impacts EBITDA through customer churn and outages, why your executive leadership team is your real "head of quality," and how automated testing serves as the ultimate guardrail to keep expensive AI agents from wrecking your codebase. Key Takeaways: The "Opinion Math" Trap: Why counting and graphing bugs to measure stability isn't science. Elisabeth reveals why these charts have zero impact on actual business outcomes and how they are easily manipulated. Quality is EBITDA: Quality is simply "value". If your portfolio company is suffering from high customer churn and frequent outages, they have a technical quality problem that is actively eroding your investment margins. The Executive Head of Quality: Real quality is determined by the systems designed by executive leadership (CTOs and VPs of Engineering), not by an isolated QA department manual-testing finished code. Keeping AI Agents Honest: AI agents have no real memory and will take shortcuts that break yesterday’s features. Robust, automated tests are the only way to keep agents honest and protect your codebase.
• 
  The 15x Fallacy: Why Coding Speed Isn't Your AI Bottleneck 17.08.2026 46mnt

  In Episode 13, Paul Karner and Dave Mangot sit down with tech veteran Michael Frendo, the current CTO of New Relic (a Francisco Partners and TPG company) and former executive at Proofpoint (a Thoma Bravo company). Michael strips away the hype of "vibe coding" and shares the hard operational realities of leading a thousand-person engineering organization through the AI revolution. He reveals why a 5x increase in coding speed only yields a 15% overall efficiency gain, why Agile has created "bad habits" that are actively hurting AI deployments, and why the true P&L winners of the next 18 months will be the firms that "skate to where the puck is going" by preparing for distributed, edge-based AI. Key Takeaways: The 15x Fallacy: Coding only accounts for 10% to 20% of the software development lifecycle. Michael breaks down why focusing solely on developer speed misses the broader delivery bottleneck—and why systemic redesign of testing, deployment, and architecture is the only way to capture true bottom-line ROI. Agile's Bad Habits: The "iterate-as-you-go" Agile culture has made engineering teams sloppy. Michael explains why AI actually demands a "go slow to move fast" approach, requiring more holistic, upfront architectural design to prevent runaway token costs and unstable systems. The 1,000-Engineer Hackathon Playbook: How Michael got an entire enterprise organization engaged in AI, generating 200 projects and 24 viable product features in a single week by shifting the responsibility of "making time" directly onto his directors. The 18-Month Horizon: Why standard LLM wrapping is a shortsighted strategy. Michael outlines how the commoditization of software delivery will shift value to small language models, localized reasoning models, and proprietary data moats.
• 
  Engineering Alpha Internally: Burning the Ships and the '30-Hour Analyst' Dividend 31.08.2026 50mnt

  How many private equity firms tell their portfolio companies to "innovate with AI" while their own deal teams are still drowning in manual Excel sheets and copy-pasting CRM data? In Episode 14 of Engineering Alpha in Private Equity, Jarrad Berman, Partner in TZP Group’s Portfolio Growth Group, joins Paul Karner and Dave Mangot to turn the operational excellence playbook on itself. Jarrad shares the story of how TZP Group — a lower middle market firm with over $2 billion in AUM—completely "burned their ships," ditched an expensive six-figure CRM, and built their own custom CRM in just two weeks using Claude. The conversation moves beyond the usual hype to dive into the hard operational mechanics of internal AI adoption. Jarrad explains how they built an interactive QSR rollup map that bypassed months of traditional sourcing to win a founder's trust in a live meeting, and shares why saving 30 hours on LBO modeling is a massive P&L win — not because you cut headcount, but because you unlock your team's analytical leverage. Key Takeaways: - The R and M Framework: How TZP Group maps every single AI initiative to a clear financial driver. It must either drive top-line Revenue (R, such as accelerating speed-to-lead from 3 minutes to 30 seconds to boost conversions) or optimize Cost and Margin (M, such as container-load optimization). - The "Burn the Ships" CRM Strategy: TZP Group was paying over six figures for a hosted cloud database that acted as a flat CRM. With a hard 30-day contract renewal deadline, they exported everything to Excel and built a custom "TZP CRM" in two weeks. The result? Annual costs plummeted from six figures to just $125 a month ($1,500/year run rate), while delivering a tool 4x to 5x more effective and custom-built for their deal workflows. - Sourcing Sourcing Sourcing (The QSR Map): Discover how TZP's deal team scraped franchise databases to build an interactive HTML target map. By identifying target coordinates and LLC ownerships, they presented a seller with 5 priority add-on locations, only to discover the operator was already secretly in contract on 3 of them—building instant, undeniable deal-table credibility. - The "30-Hour Analyst" Dividend: Shifting LBO modeling from Excel to a custom React/HTML template with slider toggles reduced a 3-day process to a few hours. Instead of "mission accomplished," this efficiency allows analysts to spend those 30 saved hours stress-testing unobvious structural variables—like interest rate spikes and tariff events—bringing massive analytical leverage and credibility to the investment committee.
• 
  The Developer Productivity Viper Pit: Why DevEx is Your Real EBITDA Driver 09.09.2026 21mnt

  When Private Equity boards want to optimize engineering costs, they often fall into a dangerous trap: using software to measure "developer productivity" using lines of code, velocity points, or pull request wait times. In Episode 15 Paul and Dave navigate the "viper pit" of software measurement. They break down why comparing velocity across teams is statistically meaningless, why vanity metrics invite gaming, and why automated productivity tools completely ignore essential "glue work"—the cross-team coordination and architectural alignment that keeps complex software systems standing. Drawing on Dr. Nicole Forsgren and Abbi Noda’s research in Frictionless as well as W. Edwards Deming’s classic management principles, Paul and Dave reveal why Operating Partners should stop chasing developer productivity and start optimizing Developer Experience (DevEx). By eliminating Lean waste — such as paying $180k engineers to sit around waiting 3 days for manual testing — operating teams can unlock massive financial leverage and drive sustainable EBITDA expansion. Key Takeaways: - Goldratt’s Law in Tech: "Tell me how you'll measure me, and I'll tell you how I will behave." Measuring developers by lines of code or story points only incentivizes teams to build bloated, unnecessary features faster. - The "Glue Work" Blindspot: As highlighted in Tanya Reilly’s seminal essay and talk Being Glue (https://www.noidea.dog/glue), the most valuable engineers on a team are often doing "glue work"—coordinating API contracts, establishing architecture standards, and aligning cross-functional teams. Vanity productivity tools rate these engineers poorly because they write fewer lines of code, creating toxic promotion incentives. - The Cost of Waiting (Lean Waste): Paying a $150k–$180k software engineer to wait 3 days for manual QA or deployment approvals is pure operational waste. Eliminating 3 days of waiting per developer across a portfolio company directly impacts the bottom line. - Deming’s Point #5 for PE: "If we improve the system, we reduce costs." Developer productivity shouldn't be measured in isolation; it must exist in service of building better products that customers eagerly pay for, with business metrics anchored right alongside engineering KPIs on the boardroom dashboard.
• 
  Meta's Experiment: Why You Can’t Outsource System Ownership to AI 14.09.2026 10mnt

  When Private Equity boards attempt to capture AI returns, a common line item targeted is engineering headcount. In Episode 16, Paul and Dave analyze a recent LeadDev report detailing Meta’s attempt to shrink engineering teams around AI agents. While Meta saw a staggering 220% increase in code commits, only 36% made it into production, while technical incidents jumped 40% and time spent firefighting production outages exploded by 70%. Corroborated by broad industry data from DX showing a precipitous drop in Developer Experience (DevEx) scores, Dave and Paul break down why "token-maxing" without human system ownership creates massive P&L liabilities. This episode outlines why "two-pizza teams" (5 to 8 engineers) remain the optimal operational unit, why replacing developers with agents starves your system of critical operational memory, and how outage firefighting actively destroys EBITDA. Key Takeaways: - The 220% vs. 36% Gap: Meta's internal experiment proved that generating code volume with AI agents does not translate to revenue-generating features. Commits skyrocketed 220%, but actual production shipping only grew 36%. - The 70% Firefighting Tax: Replacing developer oversight with autonomous agents caused technical incidents to spike 40% and developer firefighting time to surge 70%. Every hour spent firefighting outages in production is paid labor lost to reactive crisis management rather than value creation. - System Ownership Cannot Be Outsourced: You cannot outsource the operational responsibility of running a company to AI models trained on generic public code. When complex production systems break, human engineers who understand the architecture must be available to fix them. - The Return of the Two-Pizza Team: Meta’s research concluded that the ideal team size remains 5 to 8 engineers. Shrinking teams below this threshold creates severe on-call burnout, rotation failures, and institutional risk. https://leaddev.com/ai/meta-tried-to-shrink-engineering-teams-around-ai