Guest: Clement Pang - serial entrepreneur, co-founder of Wavefront (acq. VMware), former Google engineer (fraud & abuse infra), GLMX founding engineer, founder at Lavanna Labs, and now Entrepreneur in Residence at Sutter Hill Ventures.
Hello Builders!
Today’s edition of The PMF Playbook is from my podcast episode with Clement Pang. In this edition, we will cover the following topics:
● AI changes what you build, how you build, and who can build. Treat AI as a systems-level shift on par with cloud - assume competitors can move faster with smaller teams.
● Don’t be a wrapper. If your product is too close to raw model capabilities, expect the platform to ship your roadmap. Build proprietary data loops, workflows, or algorithms.
● PMF ≠ launch day. GLMX’s “Week 1: one trade” → later a $1T/day exchange. Timing + persistence + listening beats early vanity milestones.
● Performance is a feature. 10× latency wins trump long feature lists; customers rarely ask for speed but always feel it.
● Developer UX: hand the wheel over. Wavefront’’s most requested feature was a blank page + powerful query engine - users want control, not portals.
● Own (or let customers own) the data: transparency and portability without giving up operational excellence.
1) AI’s Founding Moment: Smaller Teams, Bigger Surface Area
Clement frames AI as the first industry shift since cloud:
● Ideation: Founders should use models as “thought partners” to interrogate markets, formats, sales motions, and adjacent industries.
● Build: Agents and code copilots elevate engineers; the human loop moves up to system design, interoperability, and sequencing.
● Risk: If you’re an “adjacent wrapper,” the model vendor can ship your product by next quarter. Make your defensibility live above/beside the model (workflow depth, structured data models, governance, integrations, distribution).
Hiring implication: A leaner core team can build more but beware the “single-founder + many agents” bus-factor. Institutionalize architecture, plans, and docs; don’t let your company’s knowledge live only inside one brain or a prompt history.
2) PMF in the Real World: From “Zero Trades” to a Trillion a Day
At GLMX, launch week had one trade. The lesson wasn’t “we failed” - it was “keep asking why.” Letters of intent ≠ usage; PMF requires:
● Timing: Some markets are macro-rate sensitive. Your TAM expands or contracts exogenously.
● Persistence: Iterate on the sales motion, not just the code.
● Curiosity over ego: Engineers asking business questions changes outcomes.
Playbook Prompt: After launch, schedule a standing “What did the market tell us this week?” review. No opinions - only usage, blockages, and what will unblock next usage.
3) How Wavefront Found the “Right It”
The team rapid-prototyped weekly across wild ideas, then converged on modern observability inspired by Google’s Borgmon:
● Design partners first, PMs later. Early on, founders bridged tech + product with customers (Yammer, Workday, Box…).
● Let customers feel the engine. The winning UX pattern wasn’t a dashboard portal but a query bar with power under the hood.
● Performance compounding: Quiet, deep work (adaptive schema, better tail latencies) produced the most “whoa” from customers - often more than net-new features.
Anti-Pattern to Avoid: “Zero-shot product building” via agents creates shaky foundations. Fast to first demo; slow to durability. Be willing to rewrite once you’ve validated the “right it.”
4) What to Build in the AI Era (Without Getting Sherlocked)
Clement’s de-risking heuristics:
Be more than inference. Wrap the model with data ownership, workflow state, domain-specific evaluations, and closed-loop feedback that improves over time.
Own a hard system constraint: latency, cost, scale, reliability, or a specialized retrieval/structure that models alone won’t solve.
Open Core + SaaS. Today’s observability buyer wants portability and clarity on data custody. Offer on-prem/air-gapped paths, while proving your cloud is the most efficient operator.
5) Engineering > Storytelling… or Both?
● Selling the future matters especially when AI is invisible UX. But endurance comes from IP and systems thinking: algorithms, storage formats, indexing schemes, and operating economics.
● Ask: If the foundation models got 2× better overnight, what would we still uniquely provide? Your answer is your moat.
6) The Founder’s Toolkit (Clement Edition)
A) Pre-PMF “Right It” Experiments
● Wizard of Oz tests (like IBM’s typist-behind-the-curtain): validate behavior and willingness to pay before you validate tech.
● Weekly spikes: build just enough (now with AI’s help) to check a hypothesis, then decide to double down or delete.
● Design-partner interviews that end with: “What would make this a now purchase?”
B) Build the Durable Core
● Performance budget: Treat p95/p99 latency as product requirements, not infra afterthoughts.
● Interoperability reviews: AI loves the shortest path; founders must guard long-term composability.
● Observability from day 1: Metrics, traces, structured logs - it’s cheaper than blind refactors later.
C) Moat Checklist for AI Apps
● Proprietary data pipelines with user-visible value exchange
● Domain-specific evaluation harness (precision/recall/latency/cost)
● Offline/online learning loop or feedback-to-improvement path
● Deep integrations that make you the system of action, not just of insight
● Unit economics that improve with scale (not degrade with token burn)
7) Signals You’re Close to PMF
● Users ask for less UI, more direct control (query bars, APIs).
● Your quiet releases (speed, scale, cost) trigger outsized renewal enthusiasm.
● Design partners expand usage without incentives.
● Feature requests converge; new customers ask for what existing power users already hacked together.
8) Try This Week
Map your “platform risk.” For each feature, mark: (A) likely model vendor roadmap, (B) hard integration/ops, (C) proprietary data/feedback. Prioritize B+C.
Ship a speed win. Pick one slow path users feel daily. Aim for 10× better tail latency and tell no one - watch the reaction.
Replace a portal with a prompt. Add a true power surface (query, CLI, or natural-language action that runs automations).
Data custody one-pager. Explain exactly how customers can export, self-host components, or run air-gapped. Make it a sales asset.
9) Quote Board (for your team wall)
● “AI changed what you build, how you build and whether you should build it at all.”
● “Performance is a feature. People won’t ask for it, but they’ll feel it.”
● “Give users the blank page - plus a jet engine behind it.”
● “Timing is part of PMF. Launch day isn’t verdict day.”
10) Clement’s Book Recommendation
The Right It – Alberto Savoia
Engineers love building it right; this book teaches you to build the right it - cheap experiments, demand validation, and evidence before ego.
Until next time,
Firas Sozan
Your Cloud, Data & AI Search & Venture Partner
Find me on Linkedin: https://www.linkedin.com/in/firassozan/
Personal website: https://firassozan.com/
Company website: https://www.harrisonclarke.com/
Venture capital fund: https://harrisonclarkeventures.com/
‘Inside the Silicon Mind’ podcast: https://insidethesiliconmind.com/