AI summary by TwinMind

From DevOps ‘Heart Attacks’ to AI-Powered Diagnostics With Traversal’s AI Agents

Sequoia Capital

40m 33s

Company Introduction and Vision

  • Traversal co-founders Anish and Raj building AI agents for DevOps and SRE transformation
  • Companies currently have "war rooms and armies" of DevOps/SRE engineers troubleshooting production failures
  • Production downtime costly, sometimes life-or-death situations
  • Problem worsening with AI-generated code and "vibe coding"
  • Traversal believes agents are scalable solution, seeing exciting results from deployed troubleshooter agents

Healthcare Analogy for DevOps/SRE Hierarchy

  • Stage 1: Heart attack (high severity incidents) - solve immediately, nothing else matters
  • Stage 2: Chronic conditions (daily streams of alerts, deployment safety checks)
  • Stage 3: Life optimization (5-year infrastructure planning, strategic investments)
  • Current reality: DevOps/SRE teams having "heart attack twice a week" plus chronic daily issues
  • Traversal's goal: AI handles stages 1-2, humans focus on creative strategic work (stage 3)

Impact of AI-Powered Coding Tools

  • Short-term: Tale of two worlds
    • Fast fashion coding: disposable code, reliability doesn't matter
    • Mission-critical systems: payments, financial institutions, security, streaming
  • AI tools pass local unit tests but break at scale when systems interact unexpectedly
  • Major issue: hard to debug AI-generated code due to lack of context
  • Risk: Either ban AI tools due to downtime or need new maintenance tools

Root Cause Analysis (RCA) Lifecycle

  • Customer reports issue → escalates through 5 engineering organization ladders
  • DevOps/SRE team determines incident severity/impact
  • Creates Slack incident channel with 30-50 people
  • Implicit blame game, "who done it" atmosphere
  • 10x engineer emerges as "Sherlock Holmes," figures out root cause
  • Rollback/hotfix implemented, longer-term fix planned later
  • Observability tools are #2 software spend after cloud, yet RCA remains manual

Observability Limitations

  • Current tools: MELT data (metrics, events, logs, traces)
  • Focus on data creation, storage, visualization
  • Complex troubleshooting workflow remains manual
  • Fragmented landscape: DataDog, Splunk, Dynatrace, Elastic, Grafana, ServiceNow
  • Pricing models based on data storage, no incentive for cross-platform insights
  • AI agents can automate complex workflows beyond storage/visualization

Traversal Product Architecture

  • Multi-agent system with LLM orchestration of tools
  • Tools: data fetching (logs, metrics), data processing, statistical analysis (anomaly detection)
  • Sequential approach needed - single trace may not fit LLM context
  • Gradually builds context to reach root cause
  • Mirrors human troubleshooting but more systematic, following Google SRE handbook principles
  • Less reliant on system knowledge gaps, more comprehensive coverage

Performance and Conditions

  • Works better at large enterprises vs Series A startups
  • Large enterprises: mature observability, fragmented teams, rich data available
  • Success depends on reasoning steps being findable in data
  • When root cause in data: L4 autonomous level, >90% accuracy, 2-4 minutes resolution
  • When root cause not in data: L2 level, finds important symptoms to aid human debugging
  • Gold standard evaluation: live incidents (2-3 per week), postmortem comparison

Technical Implementation

  • Offline phase: learns dependency maps using LLMs and statistics
    • LLMs understand semantic relationships between logs/tags
    • Statistics extract causal relationships from time series variations
    • Self-play prioritizes promising RCA paths
  • Online phase: uses real-time info + dependency map for incident investigation
  • Setup time: 5-10 hours (small customers), up to 1 day (large customers)
  • Read-only data access requirement

Technology Stack and Constraints

  • Must use customer's existing LLM contracts (OpenAI, Anthropic)
  • Custom models create "security hell for about a year"
  • Fine-tuning within customer environment using incident feedback
  • Architecture designed for reasoning model improvements (September bet paid off)
  • Constant 6-month technology bets required

Academic Research Translation

  • Raj's PhD: gene regulatory networks, CRISPR interventions at Broad Institute
  • Techniques for learning causal structure between genes
  • Direct application: genes → microservices, interventions → PR changes/system breaks
  • "Sheer luck" that grad school research applied perfectly

Team Composition and Culture

  • 90% engineers, mix of ML PhDs and traditional software engineers
  • Infrastructure engineers for scaling (thousands of network calls per investigation)
  • Product managers joining
  • Unified by excitement for generative AI, willingness to adapt/discard
  • Experimental mindset more important than credentials
  • Quick iteration on AI/data hypotheses

Future of Observability Teams

  • Will need fluency in both traditional systems and AI system reliability
  • AI systems break in unique ways requiring new approaches
  • Logging will fundamentally change - optimized for AI consumption vs human reading
  • More information in message fields, less concern about human readability
  • Connection between engineering systems and business logic remains challenging

Vibes Coding Future

  • Mission-critical systems will use AI tools but with better testing
  • Focus shifts from "how code was written" to "does it fulfill function"
  • Increased system interaction failures due to third-party SLA misses
  • Harder debugging without code context knowledge
  • Traversal's role becomes more critical for subtle interaction issues

Rapid Fire Responses

  • Next 12 months breakout: reasoning model applications, diagnostics, healthcare
  • Admired founder: Doina Precup (reinforcement learning researcher)
  • Recommended reading: "The Bitter Lesson" by Rich Sutton
  • Admired companies: Glean, Perplexity
  • Personal AI tools: ChatGPT exclusively
  • 5-year prediction: vibes coding in banking/healthcare "to some extent, yes"

TwinMind

Get summaries like this for your own meetings

TwinMind takes notes in the background during meetings, lectures, and calls, then writes the summary for you.

Free plan to get started

No bot joins your calls

Phone, desktop & Chrome apps

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome