YC Root AccessThe End of Manual Debugging
Aaron Epstein and Sherwood Callaway on sazabi bets AI-native observability makes manual debugging largely obsolete forever.
In this episode of YC Root Access, featuring Aaron Epstein and Sherwood Callaway, The End of Manual Debugging explores sazabi bets AI-native observability makes manual debugging largely obsolete forever Sazabi positions itself as an AI-native successor to Datadog/Sentry, letting engineers ask natural-language questions to quickly identify root causes, affected customers, and responsible commits during production incidents.
At a glance
WHAT IT’S REALLY ABOUT
Sazabi bets AI-native observability makes manual debugging largely obsolete forever
- Sazabi positions itself as an AI-native successor to Datadog/Sentry, letting engineers ask natural-language questions to quickly identify root causes, affected customers, and responsible commits during production incidents.
- The company’s central “logs are all you need” thesis argues AI makes unstructured logs newly machine-readable, reducing the need for separate metrics and traces instrumentation and simplifying observability adoption.
- Callaway’s background building infra and observability at Brex shaped his view that production is inherently unpredictable and the core job is rapid response, not perfect pre-release prevention.
- He reflects on Opkit (his first YC company in healthcare voice AI) as a misalignment with founder strengths and motivation, emphasizing the importance of building in areas of personal expertise and enjoyment.
- Returning to YC is framed as a speed and distribution strategy—imposing shipping discipline, accelerating go-to-market, and leveraging the YC network where every software company needs an observability solution.
IDEAS WORTH REMEMBERING
5 ideasAI shifts observability from manual forensics to conversational diagnosis.
Sazabi’s promise is that engineers can query production (“Why is production down?”) and get synthesized answers—turning debugging work from dashboard spelunking into guided investigation and remediation.
A logs-first approach is a deliberate bet on simplicity and adoption.
By rejecting the traditional three pillars, Sazabi argues teams can avoid heavy instrumentation overhead; logs are easiest to produce and read, and AI can extract structure and meaning that older tools couldn’t.
Observability exists because pre-production certainty is impossible.
Callaway emphasizes that tests, QA, and static analysis help, but production behavior is inherently surprising; the winning strategy is readiness and fast feedback loops once real traffic hits.
Founder-market fit can be engineered—but it’s expensive in time and morale.
Opkit shows a team can become credible in a domain over years, yet still suffer if the work doesn’t align with strengths or long-term desire; sunk cost can keep founders in the wrong game.
The bigger opportunity is maintaining software, not writing it.
He contrasts tools like Cursor (new code generation) with the ongoing burden of operating systems in production, positioning maintenance automation and “self-healing software” as the larger surface area.
WORDS WORTH SAVING
5 quotes“Logs are all you need.”
— Sherwood Callaway
“Every time there's an outage in production… I would spend hours digging through dashboards and log searches… and eventually I might get to the root cause, maybe not.”
— Sherwood Callaway
“Nothing really prepares you for production.”
— Sherwood Callaway
“We were building a futuristic AI product… and then when I would go to debug it… it was the same painful manual experience that I've had for my entire career.”
— Sherwood Callaway
“When I start this new company, it's gonna be fun.”
— Sherwood Callaway
QUESTIONS ANSWERED IN THIS EPISODE
5 questionsHow does Sazabi answer questions like “which commit is responsible?” using only logs—what correlations or signals replace traces and metrics?
Sazabi positions itself as an AI-native successor to Datadog/Sentry, letting engineers ask natural-language questions to quickly identify root causes, affected customers, and responsible commits during production incidents.
What are the failure modes of a logs-only observability system (sampling gaps, missing context, noisy text), and how does Sazabi mitigate them?
The company’s central “logs are all you need” thesis argues AI makes unstructured logs newly machine-readable, reducing the need for separate metrics and traces instrumentation and simplifying observability adoption.
Where do structured events end and “logs” begin in your worldview—are you standardizing log formats or relying on LLMs to infer structure?
Callaway’s background building infra and observability at Brex shaped his view that production is inherently unpredictable and the core job is rapid response, not perfect pre-release prevention.
If traditional metrics are “small data” and logs are “big data,” how does your custom log ingestion/storage keep costs and query latency under control?
He reflects on Opkit (his first YC company in healthcare voice AI) as a misalignment with founder strengths and motivation, emphasizing the importance of building in areas of personal expertise and enjoyment.
What is the most convincing internal or customer incident where an AI agent beat a senior engineer’s manual debugging workflow end-to-end?
Returning to YC is framed as a speed and distribution strategy—imposing shipping discipline, accelerating go-to-market, and leveraging the YC network where every software company needs an observability solution.
Chapter Breakdown
Sherwood’s second time in YC: why he’s returning now
Sherwood Callaway joins to discuss coming back to Y Combinator after exiting his first YC company. He frames this as a very different experience from his first, remote (COVID-era) batch and sets up the motivation to introduce his new company, Sazabi.
What Sazabi is: AI-native observability that answers production questions
Sherwood explains Sazabi as an AI-native observability platform built for fast-moving engineering teams—like a modern Datadog/Sentry reimagined for an AI-first world. The product goal is to replace hours of manual debugging with an interface where teams can ask direct questions about production and quickly reach root cause.
“Logs are all you need”: the manifesto and the bet against the three pillars
Sherwood presents Sazabi’s controversial principle: you can do observability well using logs alone. He argues logs are the simplest to instrument and interpret, and that modern AI makes unstructured logs far more machine-readable and useful than in the past.
Brex observability origin story: scaling microservices and getting “observability-pilled”
Sherwood recounts moving from frontend into infrastructure/DevOps and joining Brex early as one of the first infrastructure engineers. As Brex scaled to many Kubernetes microservices owned by different teams, production understanding became difficult—leading to formal observability work and a strong belief that production is fundamentally unpredictable.
What he built at Brex: auto-instrumentation, Datadog configuration, and SLO adoption
He details the practical scope of observability engineering at Brex: standardizing telemetry, building pipelines to capture/forward it, and operationalizing dashboards and monitors. He also highlights driving SLO/SLI practices so teams can consistently measure reliability and performance.
Leaving Brex to found a startup: long-held YC ambition and pandemic catalysis
Sherwood describes wanting to build a YC-backed company since his bootcamp days in San Francisco, influenced by startup culture and Hacker News. During the pandemic in New York, he and his roommate/co-founder began exploring ideas seriously, leading to their first YC application.
Opkit’s original idea: voice AI for healthcare revenue cycle workflows
Sherwood explains Opkit (YC Summer ’21) as a healthcare voice AI effort focused on automating calls to insurers. The product targeted high-friction operational tasks like eligibility checks, prior authorizations, and claim-status calls—domains with heavy phone-based processes.
Why healthcare—and the misalignment: choosing a market by “case study,” not founder-fit
He traces the rationale for healthcare: personal proximity via his father (a doctor) and the belief that verticalized fintech could be a big wave. In hindsight, he calls it a more MBA-style market selection—driven by perceived opportunity rather than deep personal insight or passion—leading to slower progress and strategic doubt.
How Opkit evolved—and unraveled: SaaS RCM, a human call center, then early LLM voice agents
Opkit shifted from RCM SaaS to leveraging LLMs for call QA/data extraction and eventually a voice agent, supported by a Philippines-based call center. Despite building early and technically challenging voice automation, fundraising traction waned, prompting a re-evaluation of whether to continue.
Exit path and next stop: joining 11x to get closer to AI product velocity
After deciding Opkit wasn’t the right long-term bet, the team explored acquirers; many felt like “more of the same” in healthcare/fintech. They chose to join 11x (an AI sales tech company) where they could work on fast-growing AI voice products with familiar connections.
The insight behind Sazabi: AI is transforming coding—but debugging is still manual
At 11x, Sherwood rebuilt a major product, then returned to on-call and maintenance realities: setting up Datadog again and debugging via the same painful workflows he’d always used. That contrast—futuristic AI development with outdated incident response—crystallized the opportunity for AI-native observability and made Sazabi feel like the “perfect” founder-market fit.
Why do YC again: acceleration, culture of shipping, and distribution to every software company
Sherwood explains returning to YC despite already being in the network: the market window feels time-sensitive and YC’s cadence can force speed and commercialization. He also wants YC’s cultural pressure toward shipping and sees YC companies as natural early customers because every software company needs observability.
Founder lessons and hiring: long-horizon integrity, and the team to build the “Linear of observability”
Sherwood emphasizes startups as a long-term compounding game where relationships matter and integrity pays off later. He closes with what Sazabi needs in hires: high-agency tool-lovers (often ex-founders), strong infra/data/storage expertise, and product/design-minded engineers to deliver a modern, elegant observability experience.
EVERY SPOKEN WORD
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
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