Qualixar film / 6:06
I Built the $6 AI Company: From Client Intake to Verified Proposal | EP 2
Can a $6 AI company take a messy client request and turn it into a verified proposal without adding another expensive SaaS stack?
In Episode 2 of The $6 AI Company, I run one client inquiry through a real open-source workflow: PDF + website + notes → structured brief → CRM handoff → idempotent automation → editable proposal → human verification.
This is not an AI-tool shopping list. Every tool has a job, every handoff is visible, and the final output stays editable.
▶ WATCH EPISODE 1 — THE $6 AI COMPANY
https://www.youtube.com/watch?v=xiBy0djq914
📖 READ THE FULL $6 AI COMPANY ARCHITECTURE
https://qualixar.com/research/blog/the-6-dollar-ai-company
📘 DOWNLOAD / READ THE IMPLEMENTATION BLUEPRINT
https://qualixar.com/learn/guides/the-6-dollar-ai-company-blueprint
WHAT YOU'LL SEE IN THIS EPISODE
• IBM Docling parsing a real client PDF into structured Markdown
• Crawl4AI extracting authorised website content with source context
• Microsoft MarkItDown converting client notes
• Twenty CRM creating a visible owner and next action
• Activepieces orchestrating the workflow and preventing duplicate submissions
• Presenton creating an editable proposal deck
• PowerPoint keeping the final deliverable human-editable
• A production mindset built around verification, idempotency and replaceable components
CHAPTERS
00:00 The Intake Pipeline & Ground Truth
00:21 The Open-Source Workflow
00:43 The Client Request
01:11 Parsing the PDF with Docling
01:36 Extracting Web Content with Crawl4AI
02:02 Converting Client Notes with MarkItDown
02:52 Twenty CRM + Activepieces Handoff
03:24 The Duplicate Submission / Idempotency Test
03:43 Production Alternatives: n8n, Windmill, pgvector
04:05 Building the Editable Proposal
04:25 Human Verification in PowerPoint
04:45 The Open-Source Reckoning
05:32 Episode 3: Stress Testing the Stack
THE CORE IDEA
AI execution is useful, but customer-facing work still needs ground truth, deterministic handoffs, visible ownership and human approval.
The goal is not maximum autonomy.
The goal is a workflow you can inspect, replay and trust.
TOOLS / TOPICS
Docling · Crawl4AI · MarkItDown · Twenty CRM · Activepieces · Presenton · PowerPoint · self-hosting · open-source AI · workflow automation · client intake automation · proposal automation · AI reliability engineering
If you are building an AI agency, consultancy, product studio or internal AI workflow, tell me which step still creates the most manual work:
Intake? Research? CRM? Automation? Proposal generation? Follow-up?
Subscribe to Qualixar for Episode 3, where we stress-test the same pipeline against harder inputs and failure conditions.
#OpenSource #AIAutomation #SelfHosted
- Published
- 2026-09-20
- Runtime
- 6:06
Connecting official YouTube player…
Press play in the official YouTube player. Playback is never started automatically.
Watch on YouTube ↗Evidence status
- Source
- Official Qualixar YouTube
- Playback
- Available here
- Search record
- Evidence complete
Read the full transcript
In the first film, we build the foundation. In this episode, we give that foundation work to do. We take one client inquiry from source material to a proposal and a follow-up draft. These are the tools we use, not as a shopping list. Dockling reads the client PDF. Crawl for AI reads an authorized website. Markdown converts notes. Twenty and Active Pieces manage the handoff. Present and PowerPoint, cal.com, and List Monk take us from proposal to review and follow up. You do not need every tool before you begin. Follow the request. Here is the request we follow. North Star Studios fictional, but the workflow is real. A workshop document, a website link, and one question. Can you help us launch it? A checked brief, a lead with a next action. An editable proposal, human approval. Follow the request. Each tool earns its place by moving that request closer closer to something a customer can use. Start with the material the client already has. Dockling converts the PDF into a form our next step can read. Check the date, opacity, and delivery format against the original source. The useful information is accessible, and we can still point to where it came from. The client also sent a website with the audience description and common questions. For this page, we use Crawl for AI. Watch the real page load, then look at the extracted markdown. The useful section is something we can compare, quote, and attach to the request. We save the address and retrieval date with it. A page can change, but a saved paragraph with its source is easy to verify. One more attachment arrives, the client's notes in a Word document. Markdown converts the file to markdown. The client's exclusions matter just as much as the sales copy. Here is the working brief. Practical workshop and launch materials included. No advertising budget, no new product development, and no unverified revenue promises. I approve the brief and lock version one. That gives you a service you can explain. Organize a client's source material and deliver a brief they can approve. Agree the scope, the review, and the handover before you quote. But useful research still needs somewhere to go. Let's make sure the inquiry gets an owner and a next action. In 20 CRM, we create the client record, attach the source inquiry, and set a visible next action, review the workshop proposal. One person, one owner, one next step. Now we connect the handoff in active pieces. When the inquiry form submits, a webhook triggers, checks the fields, and creates the lead in 20 CRM with the request ID attached. A repeated job with a visible input and a visible result. What happens when the client clicks submit a second time? In a naive setup, you get two leads, duplicate tasks, and duplicate follow-ups. Here our webhook validates the request ID. It detects the duplicate, logs the attempt, and keeps the lead count strictly at one. Safe by design. You do not have to use active pieces. If you need a vast library of pre-built integrations, choose N8N. If you want code-first workflows in Python or Windmill is built for that. For vector search, start with PGVector in Postgres before jumping to dedicated clusters. Pick the simplest tool that does the job. Now, we make the offer visible. From our approved brief, Present and Draft an editable six-slide proposal. The scope, dates, and exclusions carry over directly. No hallucinated promises, no invented deliverables. What the client agreed to in the brief is exactly what appears on the slides. A client will always ask for adjustments. Because this is a standard PowerPoint deck, a human opens the file, changes the headline to team session, and saves it. No AI regeneration loops, no loss of formatting. The deck updates cleanly, and version two is locked. You just watched one real inquiry, a PDF, a website, a Word document, move all the way to a proposal, a lead record, and a follow-up draft. Every step ran on open-source tools you can install today. No API bill, no vendor lock-in, no hallucinated promises. Eight tools, each one free, open, and replaceable. DocLink turned the brief into text, Crawl4AI read the client's website, Markdown parsed their notes, 20iCRM captured the lead, ActivePieces ran the flow idempotently, so clicking submit twice changes nothing. Present and Draft edited the deck, PowerPoint let the client edit one line. The whole run, $6 of infrastructure once. In the next episode, we take the same pipeline and stress test it. 10 more tools, one harder client, and a workflow that has to survive a bad data day. If this was useful, subscribe, leave a comment, or share this with one founder who is still doing this by hand. The next episode ships next week.
Transcript source: youtube-auto-caption. Use the film as the primary record.