Every week, hundreds of remarkable technology talks land on YouTube. Keynotes, talks, podcast episodes which is a firehose of knowledge that no one, realistically, has time to fully consume. I felt this acutely. I’d save videos to watch later, and “later” would never come.
The idea behind TLDRecap Tech was simple: what if the best insight from any tech talk could reach you in two minutes instead of two hours? So I built it. This post is about how.
🗺️ The Architecture at a Glance Link to heading
Before diving into the details, here’s the full pipeline end to end:
Every piece of this pipeline was chosen deliberately. Let me walk through each one.
🔍 The Problem: YouTube Hates Bots (Rightfully So) Link to heading
The first challenge was fetching YouTube transcripts and video metadata programically at volume. YouTube’s bot detection is sophisticated and aggressive. Residential IPs get flagged. Datacenter IPs get flagged faster. API quotas are limited for the scale I wanted.
The solution came from an unconventional place: my Android phone.
I built a custom residential proxy deployed on an Android smartphone leveraging a 5G network connection. Here’s why this works well:
- Residential IP: Mobile carrier IPs are indistinguishable from regular consumer traffic
- 5G throughput: Easily handles the bandwidth needs of metadata and transcript fetching
- Always-on: The phone sits plugged in (with charging managed by Yukti, naturally 😄)
- Rotating IPs: 5G carriers naturally rotate IPs over time, adding another layer of resilience
All transcript and metadata fetches are routed through this proxy. YouTube sees ordinary residential traffic. Problem solved no bans, no CAPTCHAs.
⚙️ The n8n Workflow: Orchestrating the Pipeline Link to heading
The entire data processing pipeline runs inside n8n, a self-hosted workflow automation platform. Think of it as the conductor of the orchestra it coordinates every step from raw video URL to a finished, publication-ready markdown recap.
The workflow steps in order:
1. Fetch YouTube Transcripts and Metadata Link to heading
The workflow starts with a video URL (or a batch of URLs from a curated list). n8n issues HTTP requests through the residential proxy to pull:
- Auto-generated or manual transcripts the raw spoken content
- Video metadata title, channel, duration, thumbnail URL, publish date, view count, and description
Raw transcripts from YouTube are noisy. They lack punctuation, speaker labels, and formatting. They go straight into the next stage for enrichment.
2. Infer Speakers Link to heading
Conference talks often involve multiple speakers, panellists, or interviewers. A flat transcript with no speaker attribution reads like a wall of text.
A Gemini model analyses the transcript and infers speaker turns based on context clues in the text phrasing shifts, topic hand-offs, self-introductions, and moderator patterns. The result is a speaker-attributed transcript that makes the final summary far more readable.
3. Summarize the Video Link to heading
This is the centrepiece of the pipeline. The attributed transcript and video metadata are passed to a Gemini model with a carefully engineered deterministic summarization prompt that requests:
- A concise
TL;DR(2-3 sentences) - Key themes and takeaways structured by topic
- Actionable insights the reader can apply immediately
- Notable quotes, where relevant
The prompt is built around determinism the same video should always produce a structurally consistent recap, regardless of which model generation handles it. This is critical when you’re balancing load across multiple models (more on that below). The output structure is locked; only model-generated content flows in.
4. URL Slug Generation Link to heading
Every recap needs a clean, SEO-friendly URL slug. Rather than slugifying the video title (which can be verbose, click-baity, or just strange), Gemini is asked to generate a concise, descriptive, lowercase slug from the video title and summary.
"Understanding Distributed Tracing in 2025 | Platform Eng Summit" becomes
/distributed-tracing-platform-engineering-2025 readable, keyword-rich, and
human-friendly.
5. Infer Tags Link to heading
Tags drive discoverability on the site and power the search index. Gemini analyses the summary and produces a set of normalized, lowercase tags covering the core topics, technologies, and domains discussed in the talk.
A talk on eBPF-based network observability might get:
ebpf, observability, networking, linux-kernel, cloud-native all
inferred automatically without any manual curation.
🤖 Gemini on the Free Tier: Round-Robin Load Balancing Link to heading
Running generative AI at scale is expensive unless you’re clever about it. Google’s Gemini series offers generous free tier limits, but those limits are per model per minute. The solution: spread the load.
I use a round-robin load balancing strategy across multiple Gemini model generations. Each incoming video is assigned to the next model in the rotation:
| Round | Model |
|---|---|
| 1 | Gemini 2.5 Flash Lite |
| 2 | Gemini 2.5 Flash |
| 3 | Gemini 3.0 Flash Lite |
| 4 | Gemini 3.1 Flash |
| … | … |
Because the summarization prompt is deterministic a consistent output schema regardless of model the recap quality remains uniform across all model generations.
This approach keeps token utilization well within free tier limits and costs exactly zero rupees in AI API fees.
🗄️ NocoDB: The Data Lake Link to heading
Every processed recap summary, slug, tags, metadata, transcript is written to NocoDB, an open-source Airtable alternative that sits on top of a relational database.
NocoDB serves as the data lake for the entire platform:
- Source of truth for all recap content
- Queryable via a clean REST API that n8n uses for writing and reading
- Visual interface for browsing, filtering, and manually overriding any AI-generated content before publication
- Audit trail every recap has its raw transcript, generated summary, and final published version stored side-by-side
The manual override capability is essential. AI summaries are good but occasionally a talk has highly specific domain context that the model mischaracterises. NocoDB makes it trivial to spot and correct these before they go live.
📝 Python + Jinja2: Rendering the Markdown Link to heading
Python script queries the API and renders the final Hugo-compatible markdown file using a predefined Jinja2 template.
The template is fixed. The AI-generated content flows into it. This separation is deliberate it means:
- Layout changes happen in the template, not in AI prompts
- Consistent front matter title, date, tags, slug, and summary are always in the right positions
- Hugo compatibility is guaranteed since the template is designed specifically around Hugo’s content model
A single script run pulls all pending recaps from NocoDB and emits a .md file
per recap into the Hugo content/ directory. From there, Hugo takes over.
🚀 Deployment: Hugo + Cloudflare Pages Link to heading
The frontend is a static site built with Hugo fast, minimal, and perfectly matched to a site that’s fundamentally a collection of markdown documents.
The deployment pipeline:
- Markdown files are committed to the GitHub repository (either manually or via an automated commit from the Python rendering script)
- The static output is deployed to Cloudflare Pages global CDN edge delivery, free tier, automatic HTTPS, and zero server management
The entire infrastructure cost of serving tldrecap.tech globally is effectively zero: free tier Gemini, free tier Cloudflare Pages, self-hosted n8n and NocoDB, and a phone I already owned.
🧩 Putting It All Together: The Full Flow Link to heading
🏁 Closing Thoughts Link to heading
TLDRecap Tech is a project built entirely at the intersection of constraints: free-tier AI, residential proxies from repurposed hardware, self-hosted open source tooling, and a static site that costs next to nothing to run globally.
The result is a pipeline that processes tech talks into polished recaps automatically, consistently, and at zero marginal cost which means I can keep the site growing without worrying about infrastructure bills scaling with content volume.
If you’ve ever felt overwhelmed by the sheer volume of great tech content out there, I hope TLDRecap Tech helps. And if you’re building something similar, I hope this breakdown saves you some of the trial and error I went through.
Check it out at tldrecap.tech and if you have feedback, there’s a form on the site. I read every submission.