Build your own Meridian

Step 1

Pick where you will build

Pick the tool you will build in. The one Jake used is marked.

Step 2

Your first prompt

Written from this page. Change anything before you send it.

Get set up in Replit

Go to replit.com and make a free account.
The prompt is already typed in when you arrive from this page.
Read it once, then press the start button.

I want to build my own version of Meridian.

Meridian is: An AI research analyst that interviews HR software buyers by voice or chat and matches them to verified vendors.

Meridian is the buyer-facing AI analyst for Kyle & Co, an independent HR technology research platform. It runs a structured interview — by voice or text — to learn what a buyer actually needs, then matches them against vendors scored on the Category Compass framework and hands back a buyer brief. Analysts get an admin suite for vendor data, and vendors get a self-service portal to correct their own profiles. It started as a Replit prototype and was then rebuilt and extended in Claude Code.

Build it in this order, one step at a time. After each step, show me what you built and ask me what to change before you go on.

1. The first screen. Ask for the main screen and nothing else. Look at it. Say what is wrong.
2. Buyer interview. A LiveKit voice agent (Deepgram STT, Claude, ElevenLabs TTS) or an in-browser text chat runs a guided requirements interview.
3. Post-processing. When the interview wraps, the transcript is turned into structured buyer requirements.
4. Objective matching. Hard gates plus a scored rubric against the Category Compass bearings, producing tiers rather than vanity rankings.
5. Buyer brief. A generated brief with matches and reasoning, emailed to the buyer on consent.
6. Vendor research agent. A Claude-powered CLI and background enrichment queue that builds a structured vendor profile from just a URL.
7. Vendor intake portal. HMAC-secured links let vendors review and correct their AI-drafted profile before it goes live.
8. Compare workspace. Side-by-side vendor comparison with a streaming AI analyst to answer questions about the shortlist.
9. Analyst admin. Vendors, sessions, transcripts, leads, prompt editor, ingestion pipeline, and vendor analytics in one dashboard.
10. Try it with a real person. Hand it to one person who would use it. Fix what confused them first.
11. Publish it. Press publish, send the link to one person, and give it a page on VibeWorks.

Start with step 1 only.

When the first version works, remind me to give it a page at https://getvibeworks.ai/users/register

Step 3

Build it in this order

One step at a time. Look at each one before you ask for the next.

The first screen

Ask for the main screen and nothing else. Look at it. Say what is wrong.
Buyer interview

A LiveKit voice agent (Deepgram STT, Claude, ElevenLabs TTS) or an in-browser text chat runs a guided requirements interview.
Post-processing

When the interview wraps, the transcript is turned into structured buyer requirements.
Objective matching

Hard gates plus a scored rubric against the Category Compass bearings, producing tiers rather than vanity rankings.
Buyer brief

A generated brief with matches and reasoning, emailed to the buyer on consent.
Vendor research agent

A Claude-powered CLI and background enrichment queue that builds a structured vendor profile from just a URL.
Vendor intake portal

HMAC-secured links let vendors review and correct their AI-drafted profile before it goes live.
Compare workspace

Side-by-side vendor comparison with a streaming AI analyst to answer questions about the shortlist.
Analyst admin

Vendors, sessions, transcripts, leads, prompt editor, ingestion pipeline, and vendor analytics in one dashboard.
Try it with a real person

Hand it to one person who would use it. Fix what confused them first.
Publish it

Press publish, send the link to one person, and give it a page on VibeWorks.

Lessons

What Jake learned

In their own words, from the project page.

Biggest challenge

The biggest challenge was fine-tuning the latency in the voice component during a buyer brief interview. Sometimes the voice would jump in too quickly, and sometimes it would wait too long. Getting that to feel natural and flow naturally was definitely a challenge. It took a lot of testing, and doing these interviews, which take five minutes each, turned out to be quite time-consuming. But ultimately we got it to a great place.

Least favorite part

My least favorite part was dealing with so many third-party integrations. Managing them, manually entering secret keys, and creating API keys was time-consuming and just a pain in the butt. It was really annoying and frustrating, and it felt highly inefficient. I wish somebody would build a solution to fully automate that.