I Built an Entire AI Content Studio for $6/Month | Zero Cloud SaaS | EP 3

Zero cloud bills, 100% open source. You already made the video. Now you need the shot, the voice over, the post, and somewhere for an interested person to go. Today, I'm turning one raw recording into that entire studio package. You'll see the finished clips, not just the tools. In our first two videos, we built the company setup and the client proposal engine. Today, we build the production studio that actually delivers the creative work for $0 in incremental software subscriptions. Here is the exact studio blueprint we are building across four chapters. In chapter 1, we take one raw workshop recording and extract our first deliverable, a verified 916 shot and an audition voiceover with omni voice. In chapter 2, we build the automated studio engine. Comfy UI for custom visual nodes, auto editor for silence removal and hyperframes for coddriven rendering. In chapter 3, we generate the full launch kit carousel platform posts and email briefs. And in chapter 4, we package the delivery, enforce client privacy, and prove our total software cost is under $6. These are all the open-source tools we are going to discuss in this video which will be the add-on to our last two videos. This is how the architecture flow will look. Let us begin with chapter one. Chapter one, the raw ingestion and short payoff. We start with one raw workshop recording. The goal is simple. Turn this into an automated client ready vertical deliverable using zero proprietary APIs. First, we run AI shots generator against the raw 4K footage. Whisper extracts timestamped speech segments while our silence detector cuts dead air automatically. The pipeline scores candidate clips for hook density, vocal clarity, and pacing. We select the strongest 18-second segment to package. The software can suggest a moment. We still decide where it begins and ends. Here is the exact vertical shot generated straight out of the pipeline. Notice the kinetic wordbyword captions, centered framing, and dynamic B-roll insertion. Produced end to end with zero manual timeline editing. A wide video does not become a good shot just because we cut off its sides. Now examine the vertical crop quality. Autoframing locks onto the speaker eyes, keeping the composition balanced without distortion or letter boxing. Crucial step, client privacy. Before any media is shared, PII sanitization removes all company secrets, usernames, and proprietary keys locally. Nothing leaves our machine. Music can wait. First, the sentence must work on its own. Now, let us replace or augment the narration. We load OmniVoice, an open-source zeros voice cloning engine running on Apple Metal. In our first two videos, we built the company setup. And in our first two videos, we built the company setup and the client pro. We do not need to make the whole voice over again. Watch how easily we fix a misspoken word, replace $14 with $6 and regenerate the audio splice seamlessly. Finally, we run automated compliance check. Every frame, caption, and audio slice is verified against client terms before handoff. The useful question is simple. Can someone tell what this offer is about? In chapter 2, we introduced the visual diffusion engine. Four tools work together. Fluke 1 for local plates, conf for node graphs, omnivoice for speech repair and hyperframes for deterministic compilation. We have the script. Now we generate the visual plate locally using eflux 1 on Apple silicon. Four steps in under 2 seconds with zero cloud API cost. Making a video shorter is useful only when it stays clear. The golden rule of AI video. Keep facts out of the diffusion latent. Watch how the generative plate stays strictly as layer zero while all typography remains live vector code. Here is the active confui node graph. Data flows from the 4-bit checkpoint loader into prompt conditioning, latent sampling, and VAE raster decoding. Listen to this comparison. Left raw studio microphone. Write zeroshot omni voice clone generated in just 4 seconds. Watch how easily we fix a misspoken word. Replace $14 with $6 and regenerate the audio splice seamlessly. Inside hyperframes, the diffusion plate binds to our kinetic DOM layout. Headless chromium compiles the entire composition at 60 frames per second. You are selling work you can deliver, not a promise of a million views. Chapter 2 is complete. Every asset, script line and visual layer is verified. Now transitioning to chapter 3, content expansion and CRM. In chapter 3, we scale from content creation to autonomous growth operations. Five open-source systems work in concert. Posters for social distribution, list monk for consent first newsletters, cal.com for inbound scheduling, 20 CRM for client deals, and active pieces for workflow automation. Postis gives us a unified console to prepare launch posts. The opening for LinkedIn explains the core problem while the caption for YouTube Shorts points directly to the result on screen. Both share the exact same offer without divergence. We've already released video one and video two. Check the cards on screen for the backstory. Interested does not always mean ready to book. Listmon provides self-hosted newsletter infrastructure. Subscribers give explicit opt-in consent. Personal data remains masked and every broadcast includes a verified one-click unsubscribe mechanism. For visitors ready to book a consultation, cal.com handles the scheduling. The interface reconciles client time in Pacific Daylight Time with host time in Greenwich Meantime, eliminating time zone confusion before an invite is sent. Here is the complete visitor journey tested end to end from video impression to landing page click, newsletter opt-in and confirmed booking. The customer never needs to understand the software stack only the value delivered. This simple agreement matters more than another clever tool. Both people should know what finished means. 20 CRM anchors the entire client relationship. The Northstar content pack deal card logs the agreed deliverables, revision limits, and an assigned team owner with explicit deadlines so nothing falls through the cracks. Small manual steps are better than automatic ones. Active pieces automates the handoff. When a deal advances to approved in 20 CRM, a web hook instantly provisions the project directory and generates delivery tasks connecting our sales pipeline directly to production. Chapter 3, growth operations are online and verified. In chapter 4, we enter production governance, tracking token costs with QAG, evaluating model quality with prompt fu, and executing the client handover. In chapter 4, we enter production governance, model evils, and client handover. Five core verification systems secure the release. Google workspace CLI for readonly delivery inspection, super local memory for decision isolation, CC usage for real cost auditing, prompt fu for assertion testing, and langfuse for trace observability. Google Workspace CLI lets an assistant inspect an approved work folder with readonly permissions. The client receives three clear structures. A final directory, editable project sources, and a short human readable readme. Keep oorthth access scoped strictly to the delivery folder. Super local memory anchors approve decisions across sessions without leaking private customer data. We store a non-sensitive note, keep the workshop date editable, and format captions for phone screens. A fresh session recalls the note with zero ambiguity. CC usage track supported local coding usage. That gives us a verified baseline, but not the whole invoice. We calculate workstation power, external API calls, hosting infrastructure, and human review time into a transparent cost sheet. This is a test mistake, not a real customer incident. Suppose a visitor asks for a refund, but our practice brief has no refund policy. This confident promise should never be sent. With Prompt FFU, we run an automated assertion to catch the hallucination before it ever reaches a client. When an AI response fails an assertion, Langfuse reveals the root cause. We inspect the exact prompt, retrieved context, and model temperature. If the assistant read an outdated document, fixing the wording alone will not prevent future errors. When a pipeline check halts, Shoutler Routes an operational alert to the developer. The notification identifies the failed task, the exact error code, and the required next action, keeping failures visible before delivery. That simple list stops us rebuilding the whole project and helps us avoid forgetting one small caption. When a client requests a date change from June 18 to June 25, we audit every dependency. Backgrounds and unvoiced B-roll are marked keep. Voiceover lines and caption timestamps are marked change. Selective rebuilding prevents accidental regressions. For client handover, we deliver an organized directory with cryptographic check sums. Anyone on the client team can find the final video, verify its integrity, and inspect editable sources without requiring a walkthrough call. Three distinct pathways emerge from one studio workflow. Creators produce viral vertical shorts. Freelancers deliver complete client content packs and founders run automated inbound growth. Start with the smallest loop you can reliably finish. That completes the $6 AI company. Head to quixar.com now for the codebase. Subscribe to our YouTube channel. Every single week we drop a production AI engineering workflow.

Transcript source: youtube-auto-caption. Use the film as the primary record.