Oct 5, 2025 · 45m · latent-space
⚡️Traversal: Causal ML and Reinforcement Learning
Anish Agarwal · 19m spoken Raaz Dwivedi · 13m spoken Alessio Fanelli · 5m spoken Shawn Wang · 2m spoken
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Traversal co-founders Anish Agarwal and Raaz Dwivedi discuss how their startup combines causal machine learning, agentic reasoning, and statistical testing to automate root cause analysis and incident remediation across complex enterprise infrastructure.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 19.1% of the talking time here. How this is scored →
The hosts as informed peer 5.5 Guest teaching 5.2 Guest disagreement 2.2 The hosts pushing back 3.3
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 32:39 Raaz rejects deterministic workflows for enterprise RCA
Raaz directly rejects the premise that deterministic workflows can replace reasoning models, explaining that enterprise gaps in instrumentation make workflows insufficient.
Hardest push from the hosts ▶ 31:57 Swyx challenges the reliance on frontier reasoning modelsSwyx pushes back against the default use of reasoning models like o3 by presenting the counter-argument that narrow debugging workflows can be explicitly scripted.
Biggest teaching moment ▶ 17:24 Founders explain statistical testing versus semantic LLM processingAnish and Raaz break down why pure LLM context dumping fails on time series data, explaining their proprietary statistical toolkit integration.
The host holds their own ▶ 22:12 Alessio dissects observability pricing and MCP incentivesAlessio leverages his hands-on developer experience with Sentry MCP to challenge the founders on enterprise pricing friction and incumbent competition.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Founders' Academic Backgrounds and the Inception of Traversal | 3 | 3 | 1 | 1 | Hosts open with a friendly welcome and ask the founders to share their background. Anish and Raaz detail their transition from academic causal ML and reinforcement learning research at MIT and Berkeley into founding Traversal. | |
| Tackling Software Maintenance Through Causal Search and AI | 5 | 5 | 2 | 3 | Alessio presses on the crowded 'AI SRE' landscape and asks what distinct approach Traversal took. Anish explains how causal ML frames troubleshooting as an adaptive search problem across massive telemetry data rather than traditional observability. | |
| Live Product Demo: Semantics Meets Statistics in RCA | 6 | 6 | 2 | 4 | Alessio role-plays a skeptical VC asking if the product is just passing telemetry into an LLM context window. The founders demonstrate how combining dynamic statistical tests with semantic agent reasoning reduces millions of tokens of time-series noise. | |
| Remediation Workflows and Market Position Against Observability Incumbents | 6 | 5 | 2 | 3 | Alessio cites his personal experience using free MCP servers versus paid observability add-ons and probes the business dynamics. Anish explains why neutrality as an outcome-based 'Switzerland' platform lets them win across fragmented enterprise toolstacks. | |
| Agentic Reasoning Models and the Importance of Rigorous Evals | 7 | 6 | 3 | 5 | Swyx brings up the contrarian viewpoint that constrained domains do not need expensive reasoning models and can rely on deterministic workflows. Raaz firmly pushes back, arguing that missing system telemetry in large enterprises necessitates world-knowledge reasoning. | |
| Future of Self-Healing, POC Strategies, and NY Team Expansion | 6 | 6 | 3 | 4 | Alessio suggests using simulated environments to speed-run customer POCs based on adjacent portfolio companies. Anish and Raaz push back with real enterprise sales experience, noting that enterprise buyers never develop conviction from staging simulations. |
Statements from this episode (13)
Agarwal: Incident troubleshooting requires adaptive search over LLM context dumping
“You cannot just Put all of it into context of an LLM and hope something great happens. You have to search the data sequentially and adaptively, right? And that's what these agentic systems are fundamentally about.”
Agarwal: AI coding assistants will cause uninterpretable outages and endless firefighting
“And so it's pretty clear to us that, you know, and we were starting to use it ourselves and sometimes we didn't understand what the code was doing, but you know, we shipped it. And so it's like, well, if this is clearly, this is going to happen a lot more. And…”
Agarwal: Sequoia met 20 AI SRE startups before backing Traversal
“Our first VC backer was Sequoia, and I think they had met, I think, like, I think like 20 VC, 20, ah, companies before this, but unbeknownst to us, and that was the first question, like Bogomol, who's a board member from there asked, like, I've heard this pitc…”
Agarwal: LLMs are really bad at processing time series data
“Because most of the data you're looking at is like time series data. And these LLMs are really bad at processing time series data, right? And that's really where like good statistics comes in.”
Dwivedi: LLMs handle semantics while statistics must handle time series
“The agent is not good at looking at time series data, so that's, that is what statistics needs to take care of. But statistics doesn't understand what is the relationship between latency and memory usage and disk utilization. So that is the LLM part.”
Agarwal: Observability incumbents lack incentive to analyze competitor telemetry data
“The typically the way they work is, is they price based on the amount of data they're storing, right? And so, you know, they have very little incentive for company A To provide you any insight on data being stored on, on company, observability company B, right…”
Agarwal: No AI incident troubleshooting competitor genuinely works in production
“And I don't think we've seen any other company in our space having something actually work at production.”
Dwivedi: Deterministic workflows cannot solve complex enterprise incident debugging
“No amount of workflows will suffice for a big enterprise. Like you have to link together some of the missing pieces, some of the poorly instrumented data, and that requires world knowledge and a few iterations with the world knowledge.”
Dwivedi: Claude is superior at agentic tool calling and error unstacking
“For some of the agentic part of the stack, we are shifting towards Anthropic because they're agentic and the tool calling, especially the unstacking part, you know, when you go down the wrong path and you build context that forces you to keep going down the wr…”
Agarwal: Rigorous eval suites are core IP for leading AI startups
“The best AI companies will always have to be the edge of what the models can do, right? I think you always want to be threading the line. If everything works all the time, then you're not really pushing the limit and you're not innovating, right? So I think yo…”
Dwivedi: AI self-healing for complex incidents is 6-12 months away
“Now for, then there is this level of 30 to 40% of the incidents or issues where you need to involve you know, a senior engineer for sanity checking. I think that healing will appear in, I don't know, six months to a year that will be comfortably, the technolog…”
Dwivedi: AI agents that reorganize whole codebases are years away
“That kind of a full agentic system that, you know, just reorganizes the whole code base. I would say, well, AI always keeps surprising you, but at least a couple of years away, if not more, where you can then rely on creating the code base.”
Dwivedi: Enterprise AI conviction requires production testing, not staging demos
“Everybody, the reason nobody gets convinced on staging environment is that everybody thinks that their environment is unique, right? Like it's the most complex thing on earth. So they're like, can your AI actually figure out the complexities of my system? Alth…”