Reads and Finds

I spend a lot of time poking at new tools as part of my own product work and investing thesis. This page is a running log of recent finds with short notes on what each one is actually betting on and where I'd expect it to break. It's not a directory and I don't take placement fees — if something made it onto this page, I spent enough time with it to have a take.

April 2026​

Land & Convert​

A daily digest of high-intent Reddit and Quora threads with AI-drafted replies. The thesis is that community engagement outperforms paid ads at the bottom of the funnel — and that the 48-hour reply window is where most of the conversion signal lives.

My take: Reddit is getting allergic to AI-sounding replies faster than models are learning to sound human, and moderator detection is the real ceiling on this category. The product only works if you use the drafts as scaffolding and write the last 30% yourself. For B2B SaaS founders with obscure ICPs, the thread discovery alone is worth the price.

Unseen Reality​

Lightweight VR glasses aimed at all-day wear rather than demo sessions.

My take: The Vision Pro taught the industry that bulk and heat kill retention more than pixel density does. The interesting question is whether a weight-first, display-second form factor can ship without compromising the thing people actually use VR for (immersion). If the first product lands even at "good enough" fidelity with genuinely wearable ergonomics, distribution unlocks itself.

March 2026​

AI for Absolute Beginners​

Plain-language guides and tutorials for non-technical readers getting started with AI.

My take: The "learn AI" content market is saturated, and differentiation almost never comes from the curriculum itself — it comes from a trusted voice and a specific audience. Worth watching if they pick a niche (e.g. AI for teachers, AI for accountants) and own that persona rather than competing on breadth.

AgentMarketCap.ai​

A leaderboard for AI agents, modeled on CoinMarketCap.

My take: Directory and ranking sites are winner-take-most, and the winner is usually whoever defines the metric everyone else cites. The hard part here isn't the UI — it's agreeing on what "usage" or "capability" means for an agent when there's no on-chain analogue of market cap. If one metric takes off (e.g. task completion rate on a shared benchmark), this becomes the default tab in every agent-builder's browser.

Clico​

A browser extension that exposes LLM actions (write, summarize, dictate) in any text field via keyboard shortcuts, using visible page context automatically.

My take: Distribution through browser extensions is one of the more underrated wedges in AI tooling right now — you don't have to convince users to switch apps. The shortcut-first UX is the right bet; copy-paste into a chat window is the friction most AI assistants still haven't removed. The risk is that OS-level AI (Apple Intelligence, Copilot in Edge) eventually eats the category from below.

SocialEcho​

Unified social media workspace for Chinese companies going global — publishing, engagement, and analytics across Western platforms.

My take: The real moat here isn't the dashboard; it's the compliance work of maintaining official API access across eight platforms whose rules change constantly. Cross-border martech for Chinese brands is a genuine underserved segment — most Western tools assume US billing and US GTM workflows.

Namefi.io​

Domain names issued and traded as NFTs on a dedicated Layer-2.

My take: Classic "bring a Web2 asset on-chain" pattern. The question I'd want answered is whether traditional domainers actually want programmable liquidity, or whether DNS already solves their problem well enough. The more interesting use case is probably developer tooling — programmable subdomains for agent identities — rather than the secondary market.

Rebyte.ai​

Visual builder for AI agents and workflows.

My take: This space (n8n, Zapier, Lindy, Flowise, and a dozen others) is getting compressed fast. The differentiator usually ends up being either deep integrations into a specific vertical or strong opinions about a specific workflow pattern — generic "build anything" builders tend to lose to the ones with a point of view.

Appifex.ai​

Idea-to-app generator.

My take: Same category as Lovable, Bolt, v0, and Cursor's agent mode. The competitive question is whether the generated code is good enough to hand off to an engineer for production, or whether it's stuck in the prototype tier. Most tools in this space are currently stuck in prototype tier and rely on users not knowing the difference.

Growbi.app​

Growth analytics with AI recommendations.

My take: Growth analytics is one of the hardest categories to start a company in — every team already has Mixpanel, Amplitude, or PostHog, and incumbents can ship AI features faster than a startup can. Most wins here come from being the default inside a specific GTM motion (e.g. PLG for devtools) rather than winning on analytics breadth.

July 2025​

BeFreed.ai​

Distills books, podcasts, and videos into summaries, audio, flashcards, and mind maps; adapts based on learning style.

My take: Blinkist-adjacent category with AI personalization on top. The real test for any learning product is 30-day retention — summaries are easy to generate, but the hard problem is getting people to come back. The 43K-user community they cite is a reasonable signal; I'd want to see time-in-app numbers before calling it.

January 2025​

Tenspect​

AI-assisted home inspection software — automated report generation and workflow management for inspectors.

My take: Vertical SaaS for the trades is one of the more under-AI'd parts of the market. Home inspectors write long, templated reports under time pressure — exactly the workload where LLMs should compress hours into minutes. The question is whether the product survives the legal scrutiny that comes with AI-generated reports in a litigious industry.

Cuckoo.Network​

Decentralized GPU network for AI inference with token-based incentives.

My take: Token-gated compute marketplaces have been attempted many times (Render, Akash, io.net). The bottleneck is almost never the hardware or the chain — it's reliability, latency, and ML-ops integration that centralized hyperscalers solved a decade ago. Worth watching if they pick a specific inference workload where cost matters more than tail latency.

Beancount.io​

Hosted frontend for Beancount, the plain-text double-entry accounting format.

My take: Full disclosure, this is a project I'm involved in. The bet is that a small but durable segment of users (engineers, IndieHackers, privacy-conscious finance nerds) prefer text files they own over a hosted SaaS ledger — and that the visualization and tax-time tooling is where they'd pay. It's a niche on purpose.