RetroLab — AI-Powered Vietnamese Tech News Platform

The Problem

Why Build RetroLab?

Vietnamese tech enthusiasts lack a high-quality, centralized source of international technology news in their native language. Existing solutions rely on manual translation, are slow to update, and provide fragmented coverage.

Content Velocity

Tech news is highly perishable. By the time human editors translate a breaking story from English to Vietnamese, the 12-hour delay renders the article obsolete and kills SEO potential.

Unsustainable Margins

Scaling a traditional localized media site requires hiring armies of translators. The human cost per article destroys profitability, making high-volume coverage financially impossible.

Inefficient Discovery

Editors waste hours manually scouring dozens of fragmented international feeds (The Verge, TechCrunch) just to decide what to write about, bottlenecking actual production.

The Solution

End-to-End AI Automation

RetroLab replaces the traditional editorial headcount with a headless, multi-agent AI architecture. By automating discovery, curation, translation, and publishing, the system operates at a scale and speed that human teams cannot match, delivering high-quality localized content at virtually zero marginal cost.

~2min Source → Published

99% Automation Rate

$0.001 Per Article Cost

Core Features

What Powers RetroLab

A deep dive into the key capabilities that make this platform tick.

01

Automated AI Pipeline with Multi-LLM Routing

To make AI-generated content profitable, I designed a cost-aware AI strategy. Routine tasks (like categorization and tagging) are dynamically routed to free or local models, reserving paid, premium APIs strictly for complex, quality-critical rewriting.

Dynamic routing between Gemini, Claude, and Ollama
Intelligent deduplication & content quality scoring
Real-time FinOps tracking to monitor per-token cost and ROI per article

admin.retrolab.com.vn/pipeline

RetroLab Pipeline Dashboard — Multi-LLM routing and task monitoring

02

Zero-Friction Headless CMS

Instead of building a custom backend UI that editors had to learn, I strategically integrated Notion as the primary database. This enterprise-grade infrastructure reduced server overhead to zero and completely eliminated onboarding friction for the editorial team.

Familiar Notion interface for zero-training onboarding
Zero database hosting costs or maintenance overhead
Seamless sync to Next.js SSR frontend with ISR caching

admin.retrolab.com.vn/editor

RetroLab Editor — Full-featured Markdown editor with split-screen preview

03

90% Workflow Reduction

The custom pipeline handles automated localization instantly. English-language source articles are culturally adapted for Vietnamese readers, transforming a multi-hour manual translation process into a single-click "Approve" workflow for human editors.

LLM-powered cultural adaptation and contextual rewriting
Automated SEO optimization (titles, tags, summaries)
Human-in-the-loop queue for final quality control

admin.retrolab.com.vn/curation

RetroLab Curation Queue — Article review from 15+ sources with approve/discard workflow

Strategic Technical Decisions

Every Tool Chosen for a Reason

Each technology was selected to solve a specific product constraint — not just because it was popular.

Sources

RSS Feeds & Web Crawlers

→

Celery Workers

Scraper → Rewriter → Image

→

LLM Engine

Gemini / Claude / Ollama

→

Notion CMS

Content Storage & API

→

Next.js

SSR Public Frontend

Why Each Tool Was Chosen

Next.js 15

Chosen for server-side rendering. Critical for SEO performance in the competitive media space.

Notion API

Chosen to eliminate editor onboarding time. Zero training needed for a tool they already use daily.

Google Gemini + Claude

Multi-provider strategy to avoid vendor lock-in and dynamically route tasks to the cheapest capable model.

Celery + Redis

Chosen for resilient async task queuing. Ensures the pipeline never blocks on slow LLM API calls.

Docker Compose

Chosen for reproducible, zero-downtime deployments on a single VPS. Keeping infrastructure costs near zero.

PostgreSQL

Chosen as the source-of-truth for pipeline state and FinOps metrics. Enabling reliable cost analytics.

Product Showcase

The Operational Engine

Behind retrolab.com.vn, the entire editorial operation runs on three interconnected systems. Each one was designed to eliminate a specific bottleneck in the content lifecycle.

admin.retrolab.com.vn/dashboard

RetroLab Dashboard — Full operational overview and human curation

Pipeline & Human Oversight

Automated Discovery with a Deliberate Quality Gate

The pipeline autonomously scrapes 15+ international RSS feeds, deduplicates content, and queues articles for processing. But full automation without oversight is reckless. The Curation Queue is a deliberate design choice: a single human decision point where an editor reviews, approves, or discards each article in one click before the AI rewriting engine takes over. This is the "human-in-the-loop" pattern that ensures editorial integrity at scale.

admin.retrolab.com.vn/finops

RetroLab FinOps Dashboard — Per-model token cost and ROI tracking

Unit Economics

FinOps — Real-Time Cost Intelligence

Proving the pipeline works is step one. Proving it's profitable is step two. The FinOps dashboard tracks per-model token spend across Gemini and Claude in real time, breaking down cost-per-article to the fraction of a cent. This is the tool that answers the single most important PM question: "At $0.001 per article, is the AI pipeline cheaper than hiring one junior editor?" The answer drives every routing decision in the system.

admin.retrolab.com.vn/editor

RetroLab Rich Content Editor — Split-screen Markdown with Notion sync

Zero-Friction Tooling

Rich Editor — Where Content Gets Polished

After an article clears the pipeline and passes curation, it arrives here. The editor provides a split-screen Markdown environment with live preview, inline AI-assisted rewriting, and one-click publish directly to Notion. By syncing with the Notion API as the content store, there is zero database hosting cost and zero onboarding friction — editors work in a tool they already know.

Impact & Results

Measurable Outcomes

Key results that demonstrate the platform's effectiveness and efficiency.

50+

Articles/Day Capacity

Automated pipeline processes and publishes 50+ articles daily from 15+ international sources without manual intervention.

~2 min

Time-to-Publish

From source article detection to fully localized, SEO-optimized Vietnamese content published on the live site.

$0.001

Cost Per Article

Multi-provider LLM routing and free-tier optimization drive per-article cost to a fraction of a cent.

99%

Automation Rate

Only the curation step requires human input. Everything else from discovery to publishing is fully automated.

Product Management

My Role & Contributions

Served as the sole Product Manager and Technical Architect, driving the platform from initial concept to live deployment.

Product Strategy

Defined the "0-to-1" product vision for a high-volume, low-overhead media platform
Defined the Vietnamese localization strategy to capture market share from slower, manual-translation competitors
Navigated resource constraints by leveraging free-tier APIs and headless integrations

End-to-End Execution

Owned the complete product lifecycle from initial architecture to live deployment
Bridged the gap between complex AI pipeline logic and a seamless CMS user experience
Architected the technical stack and managed the complete delivery cycle across frontend interfaces and backend APIs

Growth & Analytics

Designed a FinOps dashboard to aggressively track and optimize LLM token costs
Iterated on AI prompt engineering based on output quality and SEO performance
Defined success metrics and built internal dashboards to track content throughput, cost efficiency, and pipeline health

See It Live

Visit retrolab.com.vn to experience the fully automated platform firsthand.

Visit retrolab.com.vn