Tom Girgash quit his job two or three weeks before this conversation to build Taproot, a memory layer that keeps your knowledge in local files and pipes it into Claude, ChatGPT, Codex and Claude Code over MCP. Jeremy Rivera talks with him about grounding AI in your own work, finding words for a brand-new category, and building distribution without an ad budget. Follow Tom on X at @TGIRGS.

Tom took the long way into software: an Ohio State degree, a sales role selling bearings and hoses in the field, an information-systems master's at the University of Cincinnati, and a short stint as an ERP consultant at an accounting firm. Taproot started as an MVP he built for one man named Tim, who wanted one place to hold and query all his knowledge. Read the full written recap: Tom Girgash on building Taproot, the memory layer for solo AI operators.

• Why "past chats" are not memory, and why you can't tell whether what surfaced is true
• One knowledge base, many tools: why MCP makes cross-tooling the point
• The mom test, and landing on "memory layer" after months of messaging that didn't land
• About 20 signups in the first couple of weeks, Mac-only for now
• The "firm brain" idea for teams, parked while he serves solo operators first
• Pricing to a pain point: the user about to jump from the $20 tier to $100-200
• AppSumo-style launches as a paid feedback audience
• Why distribution, not code, is the hard part

Tom's open problem is the one most bootstrapped founders share: "I'm not in a position right now. I want to spend money on ads," so he needs an organic distribution engine. Reddit and forums help with validation but punish anything that smells like selling. He also names a trap of the moment: with AI you can do so much that it's easy to spend weeks tweaking a website instead of talking to users. We dig into that tension in Cracking distribution is the hard part, and Nick Eubanks makes the same case in distribution as the only moat.

Jeremy brings lessons from running product at Raven Tools: the interaction between pricing and churn, why "they left because of price" is not actionable on its own, and why a launch platform like AppSumo can buy you a committed early user base. He also maps where marketers went after the Twitter exodus: Threads, Bluesky and Slack communities.

Tom's line on where this is all heading: "I think being human, being as human as possible, connecting with your audience on a very human level, it's going to be something that's going to be pretty large going forward."

More founder stories from the show: Founder stories and bootstrapping. Try Taproot free for 30 days at taproothq.com/pricing.

Hosted by Jeremy Rivera.

When someone asks ChatGPT, Claude, Perplexity, or Gemini "what's the best X?", a few brands get named — and most companies have never checked whether they're one of them. Somya Goyal built Stellarcast to change that: a platform that watches how your brand shows up in AI answers, diagnoses why you're missing, ships the fix, and then proves the needle moved.

After 15+ years in quality engineering — eight of them at Accenture, then leading QA and building validation frameworks for AI systems — Somya is a first-time founder building Stellarcast in the open from India. In this conversation with host Jeremy Rivera, she breaks down why AI citations are a different game from search rankings, why "AI is confidently wrong" about so many brands, and how a continuous monitor-diagnose-execute-prove loop beats the one-time audit.

• The monitor → diagnose → execute → prove flywheel — and why most tools "stop at the dashboard"
• Why citations aren't rankings: AI answers overlap Google's top ten by only ~14%
• AEO in plain English — make your brand easy for AI to find, understand, and read
• Why "AI is confidently wrong", citing two-year-old pricing and reviews even after fresh data lands
• Treating LLM answers as a "liquid surface" — freshness, lag, and query-deserves-freshness in AI
• CMS integrations (WordPress, Shopify, Webflow) that turn diagnosed gaps into approved, shipped fixes
• Somya's leap from 15+ years in QA to first-time founder — and the exposure of shipping under your own name
• Market fit for SMBs and agencies, with the human always in the loop

• (00:00) Meet Somya Goyal & what Stellarcast does — a "system of AI records"
• (02:30) The four-agent loop: monitor, diagnose, execute, prove
• (05:45) CMS integrations & building in the open with pilot design partners
• (08:10) How the models behave differently — and why citations aren't rankings (~14% overlap)
• (12:20) Freshness, lag, and the "liquid surface" of LLM answers
• (16:05) AI as your least-trained support rep — the rogue knowledge base
• (19:15) Somya's background: fifteen years in QA to first-time founder
• (22:40) Market fit: SMBs, agencies, and the human-in-the-loop
• (25:30) Early access & pricing

"Visibility in AI is simple. Make your brand easy for AI to find, understand, and read. That is the new name of AEO, in very simple English."
— Somya Goyal"Ranking is different and citation is different. AI citations overlap Google's top-ten links by only fourteen percent."
— Somya Goyal"Most brands have never read what AI says about them, and sometimes AI is confidently wrong. It cites data that's maybe two years old — the pricing, the reviews — even after a lot of fresh data."
— Somya Goyal

• Stellarcast — AI brand-visibility platform (monitor, diagnose, execute, prove)
• Somya Goyal on LinkedIn — where she's building in the open
• When AI is "confidently wrong" the fix is filling the gaps with human-certified, expert-sourced content
• Monitoring your brand's AI footprint and entities is exactly where SEO advisory...

Otter and Fireflies do not fail at African languages. They discard them. Unknown languages get filtered out as noise or approximated to the nearest English word. Tobiloba Sulaimon, a self-taught developer from Lagos, built AuTrans to transcribe Pidgin, Yoruba, Hausa and Igbo as they were actually spoken. It started as a fix for a meeting the tools could not handle.

• The differentiator is what incumbents treat as an edge case. Not accuracy, but a product decision to drop what it does not recognise.
• Africa first is a constraint, not a slogan. Limited funding cannot cover many languages at once, so splitting attention means doing none of them well.
• Self-taught to shipping in roughly two years. YouTube, a first paid client, then a role where being the only developer forced the pace.
• He refuses features users ask for. Explicitly so the product never becomes a jack of all trades.
• First version shipped in about a week. The scope was small because the problem was his own.

• The gap the incumbents left open
• Africa first, and why it is a constraint
• Self-taught from Lagos
• Refusing the features people ask for
• Chapters and timestamps
• People, ideas and sources mentioned
• Questions this episode answers
• Go deeper

Tangible things were said and their machine did not capture that. And I thought that since I am a software developer, maybe I could build something that would solve this problem.

Mainstream tools filter unrecognised languages out, or map them onto the nearest English word. The output looks like a transcription failure. It is closer to a design choice, and it is the kind of choice a large company can make for years without anyone internally noticing.

That is the general lesson worth taking from this episode even if you never touch transcription. Find the thing your category's incumbents have classified as an acceptable failure.

We are African first. We are focusing on ourselves first, so I’m trying to solve this problem within Africa first before we move on.

This is not positioning. Limited funding cannot cover the continent's language count, so attempting breadth would mean doing all of it badly. The narrowness is what makes the quality possible, and the quality is the entire product.

Everything I know today I learnt on my own. I am a self-taught software developer.

HTML, CSS and JavaScript from YouTube after high school. A first paid client build. Then a front-end role at a game studio where being the only developer meant learning fast or shipping nothing. The first version of AuTrans took about a week.<...