Office Hours II: When to Pay for Vercel, Scraping Public Data for Paul's Gold-Pressure App, Audits as a Sales Hook, and Why Distribution Beats Building

Static sites cost nothing on Vercel. Two things move you onto a paid plan: backend or edge functions running at real volume, and a team of developers pushing collaborative deploys. Up to hundreds or a few thousand customers, Vercel is still the right default; Netlify and Railway are the alternatives. Harshit's own product ('Glued') is architecturally too heavy for Vercel and lives on AWS on a $1,000 free-credit grant - which he calls difficult to configure.

Paul listed the pressures he wants tracked - central banks, geopolitical issues, real rates and Fed policy, inflation, the US dollar, ETF flows, physical demand, mine supply - and asked for a cost-effective way to gather the latest data. Harshit's answer is scrapers with free tiers and API keys, wired in as MCP servers: Firecrawl by name, plus a second tool heard as 'XR'. The pattern: one scheduled scrape per source, normalise to JSON, and the app reads the store rather than the web. OH V picks Firecrawl again for Paul's compliance scanner and demos it live.

Paul wanted two modes - paste a URL, or sweep a category - and asked which analytics tools to use. Harshit reframed it: the tools are secondary; the work is (1) defining the business categories and building the list by hand, (2) writing ONE technical-SEO audit workflow that also takes a full-page screenshot and grades the design, (3) running it across every site in the list, and (4) producing comparative reports that become the cold-outreach hook - 'here is where you stand against the eleven competitors on your street'. The evaluation criteria are the product; everything else is a loop.

Harshit loves the underlying move - export your WhatsApp and email, feed it to Claude, and ask what you keep doing, what people keep asking, and what you never billed. As a personal routine (connect Gmail, scan the last N days, draft the invoice PDF) it is a single prompt in Claude or Codex. As a product it is a trust problem: nobody connects their inbox to an unknown tool unless it already looks like a CRM. He also floats a directory app that turns trending Instagram/TikTok content into buildable product ideas.

Financial-services data has to stay inside the organisation, so the answer is enterprise plumbing: AWS Bedrock, or vendor-hosted fine-tuned models (Harshit cites 'four or five' newly launched Microsoft enterprise models), never a personal Claude subscription or a raw API call from a laptop. On the career question - pitch internally or build your own - do both, with your own idea first and the employer's support as the second track; expect to use three or four tools together, not one.

A serial entrepreneur wants to sell fractional-CTO advisory to automotive dealers. Harshit's playbook: record every client conversation (a wearable note-taker), run the transcripts through Claude for the real problems, and only then talk tools and architecture - sometimes a prototype. You can serve one vertical deeply or one business function (marketing) across many. The value is exposure: his example is a marketing team that had never heard Supabase was a free backend. That one sentence was the engagement.

Both products are good and both have well-funded lookalikes (Legora, Harvey). The bottleneck is not the RAG - it is finding the buyer. Define the target precisely, use Perplexity to find the people and companies, aim for about twenty LinkedIn DMs a day, run several GTM experiments at once, and get the first five to ten clients before worrying about global scale. For Simran: prototype on five to ten documents in one legal sub-category, find one paying customer, and consider teaching lawyers to use existing AI tools as the cheaper way in. For everyone: GTM starts before the product - YC's team of two, or learn to sell through a channel you enjoy.