Akshay Katyal

After 7 years of entrepreneurship, I transitioned into an employee role. I was apprehensive—would my startup-honed skills translate? Turns out, they did. I became an intrapreneur. Here's how I leveraged my entrepreneurial mindset within established organizations: 🧵

Yeah, we are going one step further and killing runbooks. Each runbook is going to be a skill (or a combination of a few). We are not concerned about giving access to internal data as long as compliance is in place but even after that, our internal skills on claude perform way better than any off the shelf sre tool we have tried to use. Claude is really smart and combine that with our engineers expertise (programmed via skills) gets really far and is a much better use of time and resources. We have automated database investigations, heartbeat based incident investigations etc. Next is expanding this for all alarms and making infrastructure self healing.

Yeah, this is what we are also seeing. for it to work well, it needs access to your AWS cli, local runbooks etc, which teams are more comfortable to give to an internal tool. though, the usual things I have seen doesn't include self learning based on past feedback. are you guys also building something for that?

Datadog quoted Leo's team $2,000/month, with $800 going to ai-powered alert investigations capped at 20 prompts they're 3 engineers running 5 k8s clusters. didn't need a premium dashboard, needed help triaging alerts at 3am So they built their own sre agent on a €6.50/month vm It receives SigNoz alerts via webhook, investigates autonomously (checks pods, traces, recent merge requests), and either auto-fixes safe issues or dms the cto on slack with full context.. if a critical prod alert goes unacked, it nags every 10 minutes This is the pattern we keep seeing across signoz users teams aren't buying "ai sre" as a product, they're composing it. otel data, an open access layer (webhooks, mcp servers, cli), and a coding agent with a runbook leo built the whole thing in a day and open sourced it. full architecture blog and repo link below Curious, who else is building similar flows in their team to manager SRE workload?

I had a fun conversation with Ankit about how we’re getting incredible results from aggressively adopting AI in our R&D team. Check it out!

I wrote up some thoughts on why SaaS needs to become agent-first, and why we built a cli.

There are lot more details here: fin.ai/sales and ofcourse, that agent talked about in the video is live there!

I spent first 5 years of my career building a sales intelligence company. A lot of what I thought sales should look like was impossible back then! I am so glad that @Fin_ai is making it a reality! Very excited about this next step for @intercom!

It has been DAYS since my last podcast appearance, so here’s another one. This is a particularly fun chat with the incredible @robbyrussell on the On Rails podcast.

I’m very excited that today we're launching Fin for Sales. We've purpose built this new role to do the inbound sales job. Fin will now engage with prospects, answer their questions, and even qualify them - helping them and your business figure out if there’s really a fit. Fin is endlessly patient, and always has time for follow-on questions. It's very knowledgeable about your entire business. It will book meetings autonomously. And, it will get qualification information from your prospects, and route them accordingly - but always in a natural, conversational, and intelligent way. The first time we really saw end users having long detailed conversations with Fin, we knew we had something valuable. We had to build a lot of technology to make this work. Sales conversations are generally longer and more detailed than customer service ones, with more complex turn-taking and state tracking. They are also more two-sided: Fin needs to answer the user’s questions, but also has to gently drive the conversation - it has questions of its own to help figure out if there's a good fit, and what path the user should continue down the funnel. Of course, there needs to be deep integration with the surrounding CRM, to capture everything Fin learns, and give it the context it needs. We also decided to launch Fin for Sales with outcome based pricing - like the rest of Fin. We did extensive testing, and ended up on a $9.99 price point per qualification. That might sound like a lot. But you only pay it when a prospect is qualified, and we believe the qualification is deep and real enough to make that worthwhile. Having an outcome based price has forced us to build a real product that actually ensures this. We’re really delighted that we have had customers like @attio, Fellow, and Breathe, partnering with us from an early stage of the build of Fin for Sales. We're really excited to get to release this today, and see how it goes. You can find out more at fin . ai / sales - Fin for Sales is live on that page to answer any questions you might have! (Although it will send you to self-serve on our Early Stage program if you have absolutely no conversations and no plans to grow!)

I think the most important line here is: “Our Agent is strict. It won’t approve large PRs. If a change is too big, too complex, or too broad in scope, it flags it and requires it to be broken down. Creating a direct, positive incentive for engineers to ship smaller, more incremental, well-scoped changes.” This is driving real change in the PR quality. If you want your PR to be auto approved, ship smaller, safer changes. Same as how you’d do when humans review.

I do the same but with superpowers and it works like a charm 🧑‍🍳

Everyone knows that planning & designing a solution is the hardest part, especially with AI. I have started to begin my day with spawning plans for things I want to accomplish in a day since I am freshest then and use the rest of the day to implement them. It’s been working pretty well so far!

Be like @sbshah97 "This is his operating principle: each week, what’s slowing me down? Can I make it faster? Encode the answer into a skill. Compound the gains." ideas.fin.ai/p/the-ai-nativ…

@_swanand Wait… those plans don’t work in the terminal?

Anthropic really needs to allow API access within the Pro and Max plans, because their apps (the Claude Desktop app, in particular) are absolute dogshit for power users.

@loujaybee 100%! I am also a big fan of shifting these agents left and making it hard for people to ship bad stuff in the first place.

A key ingredient for a software factory is entropy management agents. Setting a clear architectural baseline of good, and then continually running agents to bring your codebase health back to that baseline.

When your CFO complains about this chart...

I have been a small part of the team building the platform that enabled us to 2x+. It’s been so rewarding with lots of hard problems to solve here and saying it is a cliche, but we are just getting started!

Our AI group published a novel finding from pre-training research In short: you can cut KV-cache memory in half by sharing most of the attention structure across heads + keeping small per-head differences—without hurting model quality or speed I'll explain this as best I can

This is no longer an experiment. Announcing the Fin CLI - the agent-first way of signing up to, configuring and using Fin and the Intercom helpdesk. Ask your agent to use fin dot ai slash cli to set Fin up from scratch.

This week at Intercom we hit over 19% of PRs auto-approved by our PR review agent based on Claude Code. Our ambitious goal is to get to 50+% by the end of this month. I'll spill all the details below and you decide yourself if we are out of our damned minds or onto something...

The result: 80% to 91%. Some from prompt improvements, half from finding a data pipeline bug that no amount of prompt engineering could have fixed. The lesson I keep learning: give Claude a clear problem, hard constraints, and tools to check its own work. The tools matter more than the instructions.

What made this work: Claude had the ability to verify every assumption. It could run classification calls, compare results, read raw data, and test hypotheses empirically. It caught that my stated baseline of 87% was stale (re-ran same-day: 80%). It validated the final prompt on 200 completely new sessions. No step was taken on faith.

I gave Claude a problem, constraints, and tools to verify its own work. Then I watched it run a genuine optimisation process. The setup: we classify Claude Code sessions into types (implement, investigate, data_query, etc). I hand-labeled ~50 sessions as ground truth. Opus gets 95% accuracy on them but is too expensive for production. Sonnet is cheap enough but was sitting at 80%. I wanted Sonnet at 95%. The constraint: every change must be backed by data. The tools: the claude CLI to run classification experiments, and 70k raw sessions to sample from.