Building in Public

Three weeks ago this was an idea. Today it's a live product with real-time voice tracking, AI coaching, Stripe payments, and 31 features.

Here's what I learned building AgendaPilot for the Replit Agent 4 Buildathon:

• Agent 4 is insanely good at full-stack Node.js/Express apps — it scaffolded the entire backend in one session
• The hardest part wasn't building — it was packaging and selling
• Your landing page matters more than your feature list
• Show, don't tell: let people try before they buy

This week I shipped the gamification system (meeting scores, badges, streaks), AI agenda suggestions, shareable report cards with PII scrubbing, and a complete navigation overhaul. The app went from "functional prototype" to "product you'd actually want to use."

Try it at agendapilot.ai

Most AI meeting tools use one model. AgendaPilot uses eleven. In parallel.

GPT-5.4 tends to be comprehensive but verbose. Claude Opus is precise but sometimes misses context. Gemini excels at structured data extraction. Grok adds contrarian perspectives.

When you send a meeting transcript through Council Mode, you don't get one opinion — you get consensus, unique insights, and explicit disagreements. It's like having 11 consultants review your meeting instead of one.

The models: GPT-5.4, GPT-5.4 Pro, Claude Opus 4, Claude Sonnet 4.6, Gemini 2.5 Pro, Gemini 2.5 Flash, Grok 4.20, Grok 3, Llama 4 Maverick, Perplexity Sonar Deep Research, and Perplexity Sonar Pro. All US-based, all running in parallel via OpenRouter.

This is the feature I'm most excited about, and it was surprisingly easy to build with OpenRouter on Replit.

Week 2 was humbling. The core product was working — voice detection, agenda tracking, AI summaries — but the edges were rough. Here are the bugs that kept me up:

• NaN% progress bar — division by zero when no agenda items existed. Classic.
• AI returning objects instead of strings — OpenRouter models sometimes wrap responses in JSON. Had to add defensive `item.text || item.task || String(item)` chains everywhere.
• Magic link emails going to spam — solved by switching to Resend with proper DKIM setup
• Speech recognition dying silently — Chrome's SpeechRecognition API has a habit of just... stopping. Added auto-restart logic.
• SSE connection drops — the live tracking stream would disconnect after 60 seconds on some networks. Added heartbeat pings and reconnection.
• Stripe webhook signature verification failing in dev — turns out the dev domain changes on restart. Had to make webhook verification environment-aware.

What Agent 4 got right: every time I described a bug, it fixed it correctly on the first try. The back-and-forth was minimal. It understood the full codebase context — when I said "fix the progress bar NaN," it knew exactly which file, which function, and which edge case.

Week 2 lesson: Shipping is 20% building and 80% debugging edge cases.

I started the Replit Agent 4 Buildathon with a one-sentence prompt: "Build a real-time meeting agenda tracker with voice detection, AI coaching, and Stripe payments."

Four hours later, I had:

• A complete Express/TypeScript server with PostgreSQL schema (users, meetings, action items, auth tokens)
• Magic link authentication via Resend
• Real-time Server-Sent Events for live agenda tracking
• Stripe integration with three subscription tiers
• A full HTML template with dark theme UI

The quality was production-grade. Not "demo quality" — actual production code with error handling, input validation, rate limiting, and proper SQL parameterization. I've built backends by hand that took 2 weeks and weren't as clean.

What surprised me most: Agent 4 didn't just scaffold — it made architectural decisions. It chose SSE over WebSockets for the real-time stream (simpler, sufficient for one-way updates). It used in-memory session tracking with PostgreSQL persistence. It set up proper CORS and helmet security headers without being asked.

Week 1 lesson: The bottleneck isn't writing code anymore. It's knowing what to build.

I'm a consultant who spends half my life in meetings. Every week, the same thing happens: we walk in with a 5-item agenda, someone goes off on a tangent about last quarter's numbers, and 45 minutes later we've covered 2 items. Nobody remembers who agreed to do what.

I built AgendaPilot because I was tired of being that person who has to say "can we get back on track?" every 10 minutes. Now the AI does it for me.

Built entirely with Replit Agent 4 for the 2026 Buildathon. Let's see what 17 days can produce.