Forem: Pascal Clément

How I built a 24/7 YouTube livestreaming SaaS with Spring Boot, FFmpeg & Imagen

Pascal Clément — Mon, 27 Apr 2026 13:51:36 +0000

How I built a 24/7 YouTube livestreaming SaaS with Spring Boot, FFmpeg & Imagen

Running a 24/7 YouTube stream sounds simple until you actually try it. OBS crashes. The machine sleeps. The stream drops.

I built IteraMachina to solve this — a SaaS platform where you compose scenes, build playlists, and stream to YouTube Live forever, fully managed in the cloud.

The Architecture

Angular SSR frontend — Scenes, Compositions, Runner in a 3-column layout
Spring Boot 3.x backend — SceneService, CompositionService, RenderService, StreamService
Imagen AI — generates visuals per scene from a text prompt
FFmpeg — assembles video from scenes + audio, streams via RTMP to rtmp://a.rtmp.youtube.com/live2
PostgreSQL — scenes, compositions, streams
Stripe — Basic / Pro subscription plans
Docker on Linode

The Workflow

Create scenes (image prompt + audio track)
Arrange scenes into a composition
Add compositions to a playlist (drag & drop reorder)
Hit Run — the stream starts and loops indefinitely

The trickiest part: @async + @Transactional

The render pipeline is async — FFmpeg can take minutes. The trap: Spring's @Async and @Transactional can't live on the same method or the transaction context doesn't propagate correctly.

Solution: split into three distinct methods — loadScenes() (@Transactional readOnly), setStatus() (@Transactional), renderAsync() (@async only). Took longer to debug than I'd like to admit.

What's next

Plan limits enforcement (Basic: 1 project, 3 scenes), Stripe customer portal, and community scene sharing.

Live at 👉 iteramachina.com — built under the Umbrincraft umbrella.

What would you add to a streaming SaaS like this? Curious what use cases others see.

How I use Gemini to score news virality — and why it actually works

Pascal Clément — Mon, 27 Apr 2026 13:17:47 +0000

How I use Gemini to score news virality — and why it actually works

When I built AsiafeedTech, I had a problem: 200+ Asian tech articles per day, but only 5 slots in the newsletter and 3 YouTube Shorts to produce.

Something had to decide which stories mattered. I built a viral scoring system powered by Gemini — and it turned out to be one of the most interesting engineering challenges of the project.

The naive approach (and why it fails)

My first instinct was keyword matching. If an article mentions "OpenAI" or "BYD" or "IPO" — high score. Simple, fast, wrong.

The problem: keyword matching finds popular topics, not viral stories. An article titled "BYD reports quarterly results" and "BYD just quietly shipped a car that charges in 5 minutes" both match "BYD" — but only one of them spreads.

What makes something actually go viral?

I broke it down into signals that a human editor would intuitively feel:

Novelty — is this genuinely new information, or a rehash?
Surprise factor — does it contradict expectations?
Stakes — does it affect a lot of people or a lot of money?
Specificity — concrete numbers beat vague claims every time
Western relevance — will a founder in SF actually care?

The last one is critical for AsiafeedTech. A story about a Chinese domestic policy change might be huge locally but irrelevant to my audience. Gemini needs to apply a cultural filter, not just a relevance filter.

The prompt

I pass the translated article title + summary to Gemini with a structured scoring prompt:

You are an editor for a Western tech audience (founders, investors, developers).
Score this Asian tech news article on a scale of 1-100 for viral potential.

Consider:
- Novelty and surprise factor
- Relevance to Western tech/startup ecosystem  
- Specificity (concrete numbers, named companies, tangible outcomes)
- Stakes (market size, geopolitical impact, technological leap)
- Emotional resonance (inspiring, alarming, surprising)

Return ONLY a JSON object:
{"score": 85, "reason": "one sentence"}

The reason field is gold — it's what I display in the admin panel and what Gemini uses as a forcing function to justify the score rather than guess.

Temperature matters more than you think

I run scoring at temperature: 0.2. Low temperature = consistent, repeatable scores across similar articles. Higher temperature introduced too much variance — the same article would score 45 one run and 72 the next.

For creative tasks like script generation I use temperature: 0.9. For scoring: determinism wins.

Does it actually work?

Honestly — better than I expected. The top-scored articles are consistently the ones I'd have picked manually. Stories about Chinese EV range breakthroughs, surprise AI model releases, or unexpected regulatory moves consistently outscore routine earnings reports.

The system isn't perfect. It occasionally over-scores geopolitical news that's more alarming than actionable, and under-scores slow-burn trends. But for a fully automated pipeline with zero human editors — it holds up.

What's next

I'm considering adding a feedback loop: if a YouTube Short generated from a high-scored article gets above-average views, that reinforces the scoring pattern. Essentially fine-tuning the prompt based on actual engagement data.

Building in public at 👉 asiafeedtech.com

What scoring signals would you add? Drop them in the comments.

How I Built a Fully Automated Asian Tech News Newsletter with Spring Boot + Gemini AI

Pascal Clément — Sun, 19 Apr 2026 10:33:14 +0000

How I Built a Fully Automated Asian Tech News Newsletter with Spring Boot + Gemini AI

From RSS ingestion to your inbox — zero manual work.

The Problem

Asian tech markets move fast. Chinese AI startups, semiconductor breakthroughs, EV innovations — most of it never makes it to Western tech media in real time. And when it does, it's already old news.

I wanted a daily digest of the top Asian tech stories, translated into English, delivered automatically. So I built it.

The Stack

Spring Boot 3 — backend pipeline
PostgreSQL + Flyway — storage and migrations
Google Gemini 2.5 Flash — translation, scoring, and newsletter writing
Buttondown — newsletter delivery (free plan, full API access)
Angular SSR — frontend
Docker Compose on Linode — production deployment

The Pipeline

The whole system runs automatically, twice a day at 10:00 and 17:00 Shanghai time.

Step 1 — RSS Ingestion

A Spring Boot scheduler fetches RSS feeds from ~15 Chinese and Asian tech sources (ithome, huxiu, 36kr, etc.). Each article is deduplicated by URL and stored in PostgreSQL.

@Scheduled(cron = "${app.run-morning-cron}", zone = "Asia/Shanghai")
public void morningRun() {
    pipeline.run();
}

Step 2 — Translation + Viral Scoring

Each new article gets sent to Gemini for:

Translation to English (title + description)
A viral score (0–100) based on category, topic, and potential Western interest

private static final String SYSTEM_PROMPT_SCORING = """
    You are an Asian tech news editor for a Western audience.
    Score each article from 0-100 based on viral potential.
    Consider: AI, semiconductors, EVs, robotics = high score.
    Local business news = low score.
    Return ONLY a number.
    """;

Category bonuses are applied on top — AI stories get +15, EV stories get +10, etc.

Step 3 — Newsletter Generation

A separate buttondown-generator service (standalone Spring Boot app) runs on its own schedule:

Calls GET /api/feeds/buttondown?limit=5 on the main backend — returns the top 5 unprocessed stories
Sends them to Gemini to write a newsletter in HTML format
Posts the result to Buttondown via API
Writes the Buttondown email ID back to each article in the DB

ObjectNode body = objectMapper.createObjectNode();
body.put("subject", newsletter.getTitle());
body.put("body", newsletter.getContentHtml());
body.put("status", "about_to_send");

restClient.post()
    .uri("/emails")
    .header("Authorization", "Token " + config.apiKey())
    .header("X-Buttondown-Live-Dangerously", "true")
    .contentType(MediaType.APPLICATION_JSON)
    .body(body.toString())
    .retrieve()
    .body(String.class);

Step 4 — The Flag System

Each article in the DB has flags per generator:

ALTER TABLE aft_feed ADD COLUMN buttondown_used BOOLEAN NOT NULL DEFAULT FALSE;
ALTER TABLE aft_feed ADD COLUMN youtube_used BOOLEAN NOT NULL DEFAULT FALSE;

This ensures no article is sent twice to the same channel, and each generator is fully independent. Adding a new generator (TikTok, LinkedIn, etc.) is just a new flag + a new endpoint.

The Architecture Decision — Separate Generator Services

Each generator (YouTube, Buttondown) is a separate Spring Boot app with its own Docker container, its own scheduler, and its own port.

Why not one monolith?

Independent deployment — update the newsletter generator without touching the YouTube pipeline
Independent failure — if Buttondown has an outage, YouTube keeps running
Independent scaling — the YouTube generator is CPU-heavy (FFmpeg), the newsletter generator is not
Clean separation of concerns

The main asiafeedtech-backend only exposes REST endpoints. Generators are clients.

What It Looks Like

Every morning and evening, this lands in subscribers' inboxes:

AsiafeedTech Daily — 2026-04-19

Yuanjie Technology: From 'Scammer' to A-share King
Yuanjie Technology has become China's highest-priced A-share stock, soaring 12x in a year...
This signals a shift in China's investment focus from consumer goods to high-tech hardware.
Read more →

Lessons Learned

Buttondown's "API access" on free plan is misleading. The general API is free, but the POST /emails endpoint (creating and sending posts) requires the Enterprise plan — unless you use the X-Buttondown-Live-Dangerously header, which bypasses the confirmation requirement for about_to_send status. This took a few hours to figure out.

Gemini rate limiting is real. For bulk scoring, we use a Semaphore(2) + delays + 429 detection. Flash is fast but not infinitely parallel.

Separate the @Async and @Transactional concerns in Spring Boot 3. Putting both on the same method causes proxy issues. Split into three methods: one for loading data (@Transactional(readOnly=true)), one for status updates (@Transactional), one for async work (@Async only).

What's Next

TikTok generator — same pipeline, different output format
Paid API access — expose the scored feed as a paid REST API for media companies and investors
More sources — Japan, Korea, Southeast Asia

If you want daily Asian tech news in your inbox:
👉 buttondown.com/asiafeedtech

Free. No spam. Unsubscribe anytime.

Built with Spring Boot, Gemini AI, Buttondown, and too much coffee.

AsiafeedTech — April Update: AI Image Gallery & Video Engine Optimization

Pascal Clément — Fri, 17 Apr 2026 06:27:13 +0000

AsiafeedTech — April Update: AI Image Gallery & Video Engine Optimization

A quick update on what shipped this month in the AsiafeedTech pipeline.

AI Visual Gallery

Every article now gets a unique cover image generated by Imagen 4.0. The prompt is built from the article title + category context + a fixed cinematic style directive. Images are saved to persistent storage and served via a proxy endpoint in the main backend — the browser never talks to the generator service directly.

You can browse the gallery at asiafeedtech.com/gallery.

Video Engine Optimization

The Shorts pipeline went through several rounds of improvement:

aformat normalization per stream for consistent audio sync in complex FFmpeg filter graphs
Dynamic font sizing for title cards based on character count
Whisper caption overflow fix via WrapStyle in the ASS template
Improved Imagen prompt structure — category-specific visual concepts per news category

TikTok Integration

The TikTok Content Posting API app is currently under review. Cross-posting will be automatic once approved — same flow as YouTube upload.

Stack

Angular SSR · Spring Boot · PostgreSQL · Flyway · Gemini · Imagen 4.0 · FFmpeg · Whisper · Docker · Linode

👉 asiafeedtech.com

## AsiafeedTech — April Update: AI Image Gallery & Video Engine Optimization A quick update on what shipped this month in the AsiafeedTech pipeline. ### AI Visual Gallery Every article now gets a unique cover image generated by Imagen 4.0. The prompt is

Pascal Clément — Fri, 17 Apr 2026 06:24:33 +0000

# How I built a fully automated Asian tech news pipeline with AI

Pascal Clément — Wed, 01 Apr 2026 08:13:56 +0000

Most Western developers have no idea what's happening in Asian tech. Not because nothing's happening — quite the opposite. The problem is the language barrier.

So I built AsiafeedTech: a fully automated pipeline that ingests, translates, scores, and turns Asian tech news into YouTube Shorts — without any human intervention.

Here's how it works:

The Pipeline

1. RSS Ingestion
25+ sources from China, Japan, Korea and SEA. Runs 3x daily via Spring Boot scheduler.

2. AI Translation & Filtering
Each article goes through Gemini with a carefully crafted prompt that translates, categorises, detects sponsored content, and filters non-tech articles in one call.

3. Viral Scoring
Gemini scores each article 1-10. We add category bonuses (AI +3, EV +2, Robotics +2) to surface the most relevant stories.

4. YouTube Shorts Generation
Top articles → Gemini script → Gemini TTS → Imagen 4.0 images → FFmpeg assembly → Whisper captions → YouTube upload.

The hardest part was getting consistent quality at every step. Each AI call needs precise prompting, output validation, and fallback logic.

Full writeup coming soon. Live at: https://asiafeedtech.com

We just upgraded the AsiaFeedTech content engine ⚡️

Pascal Clément — Mon, 30 Mar 2026 13:41:51 +0000

Here’s what’s new:

✅ Semantic deduplication (no more repeated news)
🎙️ Random AI voices (Kore / Charon)
🏷️ AI-generated tags
📺 Dynamic video length
🟥 ASS captions with highlight
🔔 Auto outro with subscribe

Cleaner. Smarter. More engaging.

visit us on https://www.asiafeedtech.com

Caption & Rendering Engine Upgrade

Pascal Clément — Sat, 28 Mar 2026 17:12:43 +0000

Update: Caption & Rendering Engine Upgrade

We improved the video pipeline with a focus on stability and readability:

Removed heavy zoom effects → lower memory usage
Optimized FFmpeg pipeline → fewer crashes (OOM fixes)
Introduced ASS-based captions → precise styling control
Improved contrast + font rendering for mobile

Result:
More stable rendering + significantly better readability.

new videos starts in: 8h

→ https://youtube.com/AsiaFeedTech/shorts

90% of Asian tech news never reaches the West.

Pascal Clément — Thu, 26 Mar 2026 08:42:37 +0000

90% of Asian tech news never reaches the West.
I built an AI Workflow that changes that:

📡 RSS from 25+ Asian sources
🤖 Translate → Script → Voice → Video → YouTube
Fully automated. Zero manual work.

👉 asiafeedtech.com
📺 youtube.com/@asiafeedtech

AI #AIWorkflow #TechNews #Asia #buildinpublic

# How we built a fully automated Asian tech news pipeline with AI

Pascal Clément — Mon, 23 Mar 2026 19:28:20 +0000

How we built a fully automated Asian tech news pipeline with AI

Most of the world's most interesting tech news originates in Asia — and most of it never makes it to English-speaking audiences.

We built AsiafeedTech to fix that.

What it does

Every 8 hours, our system:

Ingests 25+ RSS feeds from China, Japan, Korea and Southeast Asia
Translates and filters content using Gemini AI
Scores each article for viral potential (AI score + category weighting)
Surfaces only the best stories at asiafeedtech.com

The video pipeline

The top articles don't just become text posts — they become short-form videos, fully automated:

Gemini generates a YouTube script
Gemini TTS produces the voiceover
Pexels clips are sourced based on article context
FFmpeg assembles the 9:16 video
Whisper burns in captions
Branding overlay is applied
Upload to YouTube

The interesting engineering challenge: making a multi-step AI pipeline reliable, where quality compounds (or degrades) at each stage.

We'll be writing more about the architecture soon.

👉 https://asiafeedtech.com

ai #buildinpublic #webdev #opensource

AsiafeedTech – Asian tech news translated to English daily

Pascal Clément — Thu, 19 Mar 2026 11:54:16 +0000

I got tired of missing out on tech innovations coming out of Asia simply
because of the language barrier. So I built AsiafeedTech.

What it does

Aggregates 25+ RSS feeds from Chinese, Japanese and Korean tech publications
Translates articles daily using Google Gemini 2.5 Flash
Categorizes into AI, EV, Robotics, Startups, Hardware...

Tech Stack

Frontend: Angular 19 with SSR
Backend: Spring Boot 3.5 + PostgreSQL
AI: Gemini 2.5 Flash via REST API
Infrastructure: Docker

Interesting challenges

Parallel RSS ingestion with Spring @async (5 threads)
Prompt stored in DB – editable without redeployment
"Less like this" feedback updates the AI prompt automatically

Live: https://asiafeedtech.com

Kafka FinOps: How to Do Chargeback Reporting

Pascal Clément — Wed, 11 Mar 2026 06:55:18 +0000

If you run Kafka as shared infrastructure, you've probably faced this question at some point: who is responsible for this topic, and what does it cost us?

This is the core problem that Kafka FinOps tries to solve. In this post I'll explain what chargeback reporting means in a Kafka context, why it's hard, and how we implemented it in PartitionPilot.

What Is Chargeback Reporting?

Chargeback is a practice borrowed from cloud FinOps: instead of treating infrastructure costs as a single shared line item, you break them down by team, service, or product — and charge each one for what they actually use.

In AWS or GCP this is relatively straightforward. Cloud providers give you cost allocation tags. But Kafka has no native cost model. It doesn't know about teams, budgets, or ownership.

That's where chargeback reporting for Kafka comes in.

The Two Cost Drivers in Kafka

Before you can do chargeback, you need to understand what actually costs money in Kafka:

Storage — every message written to a topic is stored on disk until it expires (based on retention settings). A topic with a 7-day retention and high throughput can consume hundreds of gigabytes.

Traffic — every byte written to (bytes-in) and read from (bytes-out) a topic generates network traffic. On AWS MSK or Confluent Cloud, this traffic is billed directly.

Both can be measured via Prometheus JMX metrics:

kafka.log:type=Log,name=Size → storage per topic-partition
kafka.server:type=BrokerTopicMetrics,name=BytesInPerSec → inbound traffic per topic
kafka.server:type=BrokerTopicMetrics,name=BytesOutPerSec → outbound traffic per topic

The Missing Piece: Ownership

Metrics alone aren't enough for chargeback. You also need to know who owns each topic.

In most Kafka deployments, topic ownership is tribal knowledge. It lives in someone's head, in a Confluence page that's three years out of date, or nowhere at all.

For chargeback to work, you need a system that:

Tracks which team or person owns each topic
Links cost metrics to that ownership
Produces a report that finance or engineering management can actually use

How PartitionPilot Implements This

PartitionPilot connects to your Prometheus endpoint and takes periodic cost snapshots. Each snapshot captures storage and traffic per topic, stamped with a timestamp.

On top of that, it lets you assign an owner to each topic and consumer group. Ownership is stored in a PostgreSQL database alongside the cost data.

The result: a chargeback report in CSV format that looks like this:

Owner       | Topic                  | Storage (GB) | Traffic In (GB) | Traffic Out (GB) | Estimated Cost
------------+------------------------+--------------+-----------------+------------------+---------------
Team A      | orders.v2              | 12.4         | 45.2            | 180.8            | CHF 23.40
Team B      | user-events            | 8.1          | 120.3           | 360.9            | CHF 41.20
Team C      | analytics.raw          | 95.2         | 890.1           | 2670.3           | CHF 312.80

This report can be exported and shared with engineering managers or finance teams on a monthly basis.

Why This Is Harder Than It Sounds

A few things make Kafka chargeback tricky in practice:

Topics are shared. A single topic can be written to by one team and consumed by three others. Who pays for the outbound traffic — the producer or the consumers? There's no universal answer. PartitionPilot lets you assign separate ownership for producer and consumer sides.

Retention makes storage non-obvious. The cost of a topic depends not just on throughput, but on retention settings. A low-traffic topic with 30-day retention can cost more than a high-traffic topic with 1-hour retention.

Metrics need aggregation. Raw Prometheus metrics are per-broker, per-partition. You need to aggregate them per topic across all brokers to get meaningful numbers.

Getting Started

PartitionPilot is self-hosted via Docker Compose. You can start a free 30-day trial at partitionpilot.com — no credit card required.

If your team is running Kafka as shared infrastructure and you want to start doing proper cost allocation, give it a try.

Pascal Clément — founder of PartitionPilot

Forem: Pascal Clément

How I built a 24/7 YouTube livestreaming SaaS with Spring Boot, FFmpeg & Imagen

How I built a 24/7 YouTube livestreaming SaaS with Spring Boot, FFmpeg & Imagen

The Architecture

The Workflow

The trickiest part: @async + @Transactional

What's next

How I use Gemini to score news virality — and why it actually works

How I use Gemini to score news virality — and why it actually works

The naive approach (and why it fails)

What makes something actually go viral?

The prompt

Temperature matters more than you think

Does it actually work?

What's next

How I Built a Fully Automated Asian Tech News Newsletter with Spring Boot + Gemini AI

How I Built a Fully Automated Asian Tech News Newsletter with Spring Boot + Gemini AI

The Problem

The Stack

The Pipeline

Step 1 — RSS Ingestion

Step 2 — Translation + Viral Scoring

Step 3 — Newsletter Generation

Step 4 — The Flag System

The Architecture Decision — Separate Generator Services

What It Looks Like

Lessons Learned

What's Next

Subscribe

AsiafeedTech — April Update: AI Image Gallery & Video Engine Optimization

AsiafeedTech — April Update: AI Image Gallery & Video Engine Optimization

AI Visual Gallery

Video Engine Optimization

TikTok Integration

Stack

## AsiafeedTech — April Update: AI Image Gallery & Video Engine Optimization A quick update on what shipped this month in the AsiafeedTech pipeline. ### AI Visual Gallery Every article now gets a unique cover image generated by Imagen 4.0. The prompt is

# How I built a fully automated Asian tech news pipeline with AI

The Pipeline

We just upgraded the AsiaFeedTech content engine ⚡️

Caption & Rendering Engine Upgrade

90% of Asian tech news never reaches the West.

AI #AIWorkflow #TechNews #Asia #buildinpublic

# How we built a fully automated Asian tech news pipeline with AI

How we built a fully automated Asian tech news pipeline with AI

What it does

The video pipeline

ai #buildinpublic #webdev #opensource

AsiafeedTech – Asian tech news translated to English daily

What it does

Tech Stack

Interesting challenges

Kafka FinOps: How to Do Chargeback Reporting

What Is Chargeback Reporting?

The Two Cost Drivers in Kafka

The Missing Piece: Ownership

How PartitionPilot Implements This

Why This Is Harder Than It Sounds

Getting Started