CHAPTERS
- 0:00 – 1:00
AI-written code and the looming debugging/downtime crisis
Anish frames a hidden risk of AI: exploding code volume and shrinking human context, making failures harder to diagnose. He connects this to real-world outages and the enormous enterprise cost of downtime.
- •AI tools are generating far more code than teams can fully understand
- •Lack of full-system context makes troubleshooting incidents dramatically harder
- •Downtime is a massive economic problem (hundreds of billions annually)
- •Large outages illustrate how long recovery can take and how broad the impact is
- 1:00 – 1:30
Traversal’s mission: an AI Site Reliability Engineer for incident root cause
He explains what Traversal builds and how it changes the on-call workflow. The product aims to surface candidate root causes with evidence before engineers even join the incident channel.
- •Traversal provides automated incident triage and root-cause candidates
- •“Citations” point to internal observability data (logs/metrics/traces), not web links
- •Reduces the common escalation spiral where many teams join and blame-shift
- •Goal: shrink incident response from many people/hours to a few people/minutes
- 1:30 – 2:31
Company context, funding, and early customer validation (DigitalOcean)
Anish introduces himself, Traversal’s traction, and why the problem matters at cloud scale. He highlights measurable impact from working closely with DigitalOcean.
- •CEO/co-founder; raised $48M seed + Series A (Sequoia, Kleiner Perkins)
- •Worked closely with DigitalOcean for ~6 months
- •Reported ~37–40% reduction in time to resolution
- •In cloud infrastructure, outages cascade to hundreds of thousands of customers
- 2:31 – 3:01
From competition and math to MIT: early AI inspiration
He shares his personal background and why he pursued AI deeply. A key moment was seeing AlphaGo’s creativity-by-self-play as a sign of broader applicability.
- •Competitive mindset plus attraction to “clean” abstractions in math/science
- •Came to MIT (2016) to study ML/AI deeply
- •NeurIPS 2017 AlphaGo keynote as a formative spark
- •Belief that self-improving AI architectures could generalize far beyond games
- 3:01 – 4:32
The ChatGPT inflection point and the pull toward building
Anish describes ChatGPT’s arrival as a world-changing moment and contrasts research vs. entrepreneurship. He argues fast feedback loops matter even more in rapidly shifting AI landscapes.
- •ChatGPT era felt like a fundamental, once-in-a-lifetime shift
- •He’s drawn to uncertainty and “zero-to-one” creation
- •Entrepreneurship compresses feedback cycles from years to weeks/months
- •In fast-moving AI, quick iteration is a strategic advantage
- 4:32 – 5:02
Start at the edge: choosing a problem at the frontier of models
He explains how the best AI companies differentiate by operating near the limits of current models. Traversal began with a problem “shape” in mind, then searched for the right wedge.
- •Differentiation comes from building at the frontier where models are evolving
- •Being “at the edge” requires rapid detection of what works and course correction
- •Traversal started (Jan 2024) without a fixed idea but with a clear problem taste
- •Sought an intersection of causal ML, reinforcement learning, and AI agents
- 5:02 – 6:02
Why incidents were the perfect fit: causal reasoning + agents + messy data
The team converged on incident response as a needle-in-haystack problem well-suited to LLMs and agentic workflows. He outlines why the market is large and growing as code volume explodes.
- •Incident response resembles finding true causes amid many false signals
- •Haystack data includes logs, metrics, traces, code, configs, and more
- •Agentic workflows can automate sequential querying and reasoning steps
- •Market demand is strong because everyone cares about preventing/ending outages
- •AI-generated code growth makes debugging harder, increasing the need
- 6:02 – 7:03
Investor thesis and first-principles AI product thinking
Anish describes alignment with Sequoia’s view: high technical risk but low market risk if solved. He emphasizes that the next wave comes from leveraging what AI is uniquely good at—not copying human processes.
- •Sequoia shared the thesis: big market if the hard technical problem is cracked
- •AI agent companies often focus too much on mimicking humans
- •First-principles design: exploit model strengths rather than replicate workflows
- •Reinvention happens when you redesign processes around new capabilities
- 7:03 – 7:33
MVP success vs. enterprise reality: the 90% to 0% accuracy shock
He recounts early MVP wins with small companies and then the painful failure when moving to larger enterprises. The lesson: production-grade AI in complex environments is fundamentally different from demos.
- •MVP built quickly (June/July) and worked well on smaller-scale data
- •Early approach used historical incidents and playbooks to drive an agent
- •At enterprise scale (e.g., DigitalOcean), accuracy collapsed to ~0%
- •Key distinction: MVP performance doesn’t translate to production complexity
- 7:33 – 8:04
The first principle that unlocked reliability: use computation, not “creativity”
Traversal re-architected the system to align with what models do best—systematic inference over evidence. This shift restored performance and positioned the product to improve with stronger reasoning models.
- •Rebuilt architecture to rely on computation/inference instead of handcrafted cleverness
- •Accuracy returned to ~90% after the redesign
- •Critical question: how to get better as new reasoning models arrive
- •Many competitors didn’t improve with reasoning models—architecture mattered
- 8:04 – 9:04
Reasoning models as detectives: connecting evidence to a single root cause
He explains why modern reasoning models excel at “mystery solving” with many moving pieces. Incident diagnosis matches this pattern: many symptoms, one underlying cause, and a need for evidence-backed conclusions.
- •Reasoning models are strong at detective-story style evidence synthesis
- •They connect disparate clues to arrive at a clear final answer
- •Incidents present multiple symptoms across systems that must be reconciled
- •Maximizing model strengths is key to durable advantage
- 9:04 – 12:57
How to survive the AI coding era: reinvent maintenance and choose your people
Anish returns to the hidden risk: context loss as AI writes more code, pushing engineers toward endless troubleshooting unless maintenance is reinvented. He closes with lessons on resilience, recovery, and building with people you respect and enjoy.
- •AI-generated code increases complexity and reduces shared context for debugging
- •Engineers risk shifting from creative design to constant troubleshooting
- •Software maintenance must be rebuilt alongside AI-driven software creation
- •Resilience and grit matter when outcomes swing between high highs and low lows
- •Long-term success depends on surrounding yourself with mentors, partners, and teammates you truly value
