How I AIHow Amplitude built an internal AI tool that the whole company’s obsessed with (and how you can too)
CHAPTERS
- 0:00 – 2:53
Why Moda went viral inside Amplitude: unlocking enterprise data for everyone
Claire Vo introduces Wade Chambers (Amplitude’s Chief Engineering Officer) and tees up Moda—an internal AI agent that quickly spread across the company. The framing is that internal AI tools can compress workflows like answering business questions and generating product artifacts (e.g., PRDs).
- 2:53 – 4:55
Build vs. buy: why Amplitude built an internal AI tool in weeks
Wade explains the decision-making behind building instead of buying and emphasizes the speed/low-risk nature of internal AI prototypes. He highlights leveraging off-the-shelf components (like Glean APIs) to avoid overengineering while still unlocking internal data access.
- 4:55 – 7:19
What Moda is: an internal agent for Q&A, research, and artifact generation
Moda is positioned as a company-wide interface to internal knowledge that supports answering questions and producing work outputs. The north star is becoming “AI-native” by making internal data broadly accessible and actionable.
- 7:19 – 9:17
Adoption by design: the “social engineering” strategy (build it where people work)
Wade explains the adoption thesis: meet employees where they already collaborate (Slack), keep usage visible, and let people learn from each other’s prompts and outputs. Seeing credible peers use the tool creates trust and accelerates spread.
- 9:17 – 10:58
Slack demo: Moda introduces itself and emphasizes citation/verification
Wade demonstrates Moda in Slack, showing how it responds quickly and describes its capabilities. A key element is that it cites sources so employees can verify answers rather than blindly trusting generated content.
- 10:58 – 12:43
What data Moda can access (and what it intentionally can’t)
The conversation enumerates Moda’s connected enterprise sources and clarifies the access model. The tool focuses on “enterprise-public” content and avoids private/personal/restricted datasets, reinforcing trust and governance boundaries.
- 12:43 – 17:41
Company-wide usage analytics: using Moda to product-manage Moda
Wade shows how Moda can be queried to understand who uses it and what questions are being asked, turning the tool into a feedback loop for its own improvement. This helps identify friction points, missing data access, or prompt/UX issues.
- 17:41 – 23:24
Product insight workflow: theme mining from customer feedback across systems
They walk through a common PM workflow: start broad, identify themes, then drill into a promising subtheme backed by real customer quotes and evidence. Moda pulls from tools like Zendesk/Productboard/call transcripts to surface where there’s “heat.”
- 23:24 – 27:30
Behind the scenes architecture: Slack + web UI on top of a framework and RAG search
Wade explains the technical setup: a Slack bot and a web UI call into an internal processing framework (Langley), and Glean APIs are used to retrieve relevant enterprise content for RAG-style grounding. The system supports both general chat and specialized workflows (like PRD generation).
- 27:30 – 29:09
Prompt/agent quality: how they learned to write strong prompts and iterate
Claire probes how the team got good at structured prompts and multi-step orchestration. Wade says part was existing expertise, and part was using AI recursively to improve prompts, then editing and operationalizing them in version control.
- 29:09 – 32:37
From one customer insight to a multi-step PRD (plus prototype prompts)
They demo Moda’s PRD flow: starting from a single-sentence insight, it generates a full PRD including problem exploration, solution exploration, requirements, and prototype-generation prompts. Outputs are visible to others, enabling reuse and review.
- 32:37 – 34:38
Do teams actually use AI PRDs? Review loops, variants, and critical thinking
Claire challenges whether people read AI-generated docs or skip straight to prototypes. Wade says they still review problem/solution/requirements, look for cons and counter-evidence, and explore multiple prototype variants—treating AI as acceleration, not autopilot.
- 34:38 – 36:00
Organizational impact: faster product cycles and role-swapping across PM/Design/Eng
Wade describes a velocity shift: research → PRD → design → build can be compressed into a single working session, especially for scoped initiatives. They even run intentional “role swap” exercises where designers code and engineers write requirements to build fluency and empathy.
- 36:00 – 40:26
Lightning round: AI for engineering productivity, tech debt, and ‘debugging’ prompts
Wade shares what excites him about AI-assisted coding—especially tackling tech debt and reducing toil—while acknowledging the need to make systems more AI-friendly. He also gives practical strategies for when an LLM isn’t behaving: rewind to the last good step and provide precise feedback.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome