Five systems, one idea.

The problem

Making a researched, scripted, voiced and edited video is a long chain of manual steps. Each one waits on a person. Quality drifts from video to video. The bottleneck is never the work itself — it's the waiting between steps.

How it works

1. A topic is submitted through a form and a tracked job is created.
2. The input is cleaned, and the niche is decided automatically.
3. Sources are searched and quality-checked, and claims are verified against them.
4. A script is written from the verified claims, then a separate critic stage approves it, sends it back, or patches specific lines.
5. Voiceover, captions, music and sound effects are produced and mixed.
6. An AI model plans the shots, matching stock footage is collected, and the video is rendered.
7. A final automated check decides whether the video passes.

Built with

• n8n
• Supabase
• Google Gemini
• NVIDIA Nemotron
• Local TTS

Design decisions

• A quality gate after each major stage
• Retries with adjusted input before anything is escalated
• Every run logged, so each video can be traced
• A kill switch that stops the pipeline cleanly

The problem

Finding good prospects means searching, opening websites, judging each business, finding a contact, and writing a tailored email. It eats hours before a single conversation starts. Most of it is the same thing over and over.

How it works

1. Search queries are built from one configuration step.
2. Junk and directory pages are filtered out, and duplicates are removed, including leads found in earlier runs.
3. Each homepage is fetched, and pages with too little content are dropped.
4. A local AI model analyses the business, finds weaknesses, and writes a short opportunity summary.
5. The lead is scored, and only qualified leads continue.
6. A contact is found and a personalised draft is written.
7. Everything is saved as one organised row per lead.

Built with

• n8n
• Ollama (local models)
• Google Sheets
• FastAPI

Guardrails

• Caps on queries and leads per run
• Protection against instructions hidden in scraped pages
• Automatic retries and rate-limit waits
• Nothing is ever sent without a person's approval

The idea

A portfolio that only changes when its owner remembers to update it goes stale. This engine removes that dependency. You submit an idea, the system drafts it, and you decide whether it goes live — or the system publishes it for you after 48 hours if you do nothing.

How it works

1. You send a short submission from Slack — title, type, and a few accomplishment bullets.
2. A drafting stage writes the post in the right shape: case study, lesson learned, or blog post.
3. A critic stage scores the draft, verifies the claims against what you submitted, and strips anything confidential.
4. The draft lands in Supabase with a publish time set 48 hours out.
5. A Slack card arrives with four buttons: Approve, Reject, Hold 7 days, Open in Notion.
6. Approve publishes now. Reject buries it. Do nothing, and it publishes itself at 48 hours.
7. The Notion portfolio mirrors every published post automatically, on its own clock.

Built with

• n8n
• Supabase
• Notion API
• Slack Bolt
• Ollama

Design decisions

• Supabase is the single source of truth. Notion is a mirror, not a co-author.
• The 48-hour timer lives in the database, so auto-publish works even when n8n is down.
• One writer per field, so there are no sync loops to break.
• A confidence score, not pass/fail, so a broken critic is visible instead of silent.

The problem

Monitoring a marketplace for pricing anomalies and genuine seller trust is impossible to do manually. Bait prices, troll listings, and untrusted sellers distort the market, making it hard to price items accurately or safely.

How it works

1. Fetch Listings: Rotates through pet catalog searches, fetching up to 400 listings per run.
2. Normalize: Uses regex to extract canonical item names, drops junk, and handles price conversions (cents to dollars).
3. Seller Profiling: Fetches Gameflip profile data (feedback, account age, ID verification) with a 6-hour cache and TOTP authentication.
4. Market Analysis: Aggregates listings by item and tier (Regular, FR, NFR, MFR), filtering out junk and troll titles.
5. Trust Gate: Rejects sellers with low feedback percentages, too few reviews, or new accounts.
6. Statistical Trimming: Uses median-based filtering to remove low-bait and high-troll outliers.
7. Price Calculation: Applies dynamic markups (10% or 13%) based on market value and validates against historical anchors.

Built with

• n8n
• PostgreSQL
• JavaScript
• Gameflip API
• TOTP Auth

Design decisions

• Cached seller profiles to avoid API rate limits.
• One price per seller to prevent market skewing.
• Strict tier inversion checks (a higher tier can never be cheaper than a lower tier).
• Automated alert logging for suspicious price jumps.

The problem

Creating content for multiple platforms means writing captions, tags, and blog posts for each one manually. It's repetitive, time-consuming, and leads to inconsistency across channels.

How it works

1. Discovery: Watches Google Drive folders for new video files.
2. Transcription: Extracts audio with FFmpeg and transcribes locally using Whisper.
3. Caption Generation: Uses Groq (GPT-OSS) to generate platform-specific captions and hashtags (YouTube, TikTok, IG, FB).
4. YouTube Scheduling: Uploads video, sets thumbnail, and schedules natively.
5. Facebook Scheduling: Uploads video and schedules natively to the correct Page.
6. Blog Generation: NVIDIA Nemotron writes a full blog post from the transcript.
7. Image Generation: Cloudflare Flux creates cinematic images for the blog.
8. SEO Gate: Validates word count, headings, keywords, and banned words.
9. Blogger Scheduling: Creates a draft and schedules it, embedding the YouTube video.

Built with

• n8n
• Groq
• NVIDIA Nemotron
• Cloudflare AI
• Blogger API
• YouTube API
• Facebook Graph API
• Whisper
• FFmpeg

Design decisions

• Human-in-the-loop via Google Sheets (statuses are tracked, errors logged).
• Strict SEO gate prevents AI fluff from being published.
• Dynamic tags and hashtags generated from transcript keywords.
• Native scheduling APIs used for each platform to ensure maximum reach.