Thursday Takeover

A
Ashvin Praveen·

Every Thursday a group of builders meet up at the 500 Global office and share what they're tinkering with & learning. Here's some of the key learnings this week:

The Momentum Trap*: Why it's harder to kill a "successful" misaligned project than to start a new one.

Localized AI*: Generic models fail at local nuance—there's a massive gap for region-specific context (Malay/Chinese).

Newsroom Automation*: How AI is collapsing the 50-person editorial workflow into a single dashboard.

Privacy-First Architecture*: Why moving to Zero Data Retention (ZDR) is becoming non-negotiable for serious AI tools.


1. Mohtasham: Localized AI Chat

🔍 Overview

  • Project: chat.mohtasham.dev

  • Core: Interface to chat with multiple models (including YTL's Ilmu).

  • Strength: Superior performance in Malay and Chinese language contexts.

🎬 Reels Script

Hook: "Your LLM doesn't speak 'Malaysian'—and that's a problem." Shotlist:

  • [0:00] Selfie style, pointing at a screen showing standard ChatGPT failing a local slang query.

  • [0:03] Screen recording: Switching to Mohtasham's interface.

  • [0:05] Typing a complex query in mixed Malay/English (Manglish).

  • [0:08] The model outputting a perfect, culturally nuanced response.

  • [0:12] Cut to face: "We need models that understand context, not just language."

Caption: Generic AI is great. Localized AI is game-changing. 🇲🇾 #AI #LocalLLM #TechMalaysia

📝 Blog Post

Title: Why Context is King: The Case for Region-Specific LLMs

We treat Large Language Models like universal translators, but they have a blind spot: Culture.

If you ask GPT-4 to write a formal email, it excels. But if you ask it to explain a concept using Malaysian slang, or to navigate the complex linguistic dance of "Manglish" (a blend of Malay, English, Chinese, and Tamil dialects), it stumbles. It translates the words, but it misses the soul.

This week, Mohtasham demoed chat.mohtasham.dev, and it highlighted a massive gap in the current AI landscape. By integrating models like YTL's Ilmu, which are trained specifically on local datasets, the difference in output isn't just grammatical—it's cultural.

The "Global" Problem Silicon Valley builds models for the world, but "the world" usually means "Western context." When these models encounter high-context Asian languages or dialects, they default to sterile, textbook translations. They lack the "grounding" of someone who actually lives here.

The Local Solution Mohtasham’s interface allows users to toggle between generic giants and hyper-local specialists. The result? A model that understands that "Can lah" acts as a full sentence, and that the tone of a message changes entirely based on which particle you use at the end of it.

For builders in Southeast Asia, the opportunity isn't to build another wrapper around OpenAI. It’s to build the layer that actually understands us.

💼 LinkedIn Post

Most people think "Translation" is enough for AI. It isn't.

Translation converts words. Localization converts meaning.

Mohtasham showed us chat.mohtasham.dev this week, and the difference is stark. When you chat with a model trained on local context (like YTL's Ilmu), it doesn't just get the grammar right—it gets the vibe.

For builders in SE Asia: The opportunity isn't just wrapping GPT-4. It's building the layer that actually understands us.


2. Vivian: The Momentum Trap (Cermini AI)

🔍 Overview

  • Project: Cermini AI (Brand identity via LLM).

  • Status: Shut down despite traction, funding interest, and partners.

  • Lesson: "The Momentum Trap"—harder to stop a misaligned success than start a new failure.

🎬 Reels Script

Hook: "We shut down our startup. And it was succeeding." Shotlist:

  • [0:00] Text on screen: "Traction? ✅ Investors? ✅ Quit? ✅"

  • [0:02] Montage of team working, graphs going up.

  • [0:05] Vivian speaking (or voiceover): "We had the partners, we had the government backing."

  • [0:08] Sudden cut to black/pause: "But we didn't have the excitement."

  • [0:10] Selfie style: "Don't let momentum drag you down the wrong path."

Caption: The hardest decision isn't starting. It's stopping when it works but feels wrong. 🛑 #FounderStories #Pivot #StartupLife

📝 Blog Post

Title: The Momentum Trap: Why We Killed Cermini AI

There is a dangerous misconception in the startup world that the hardest thing to do is fail. It isn't. The hardest thing to do is succeed at the wrong thing.

Vivian shared a story this week that sent a shiver down the spine of every founder in the room. Her project, Cermini AI, was working. By every traditional metric, it was a "green light."

  • Market Validation? Check. Users loved the brand identity insights.

  • Partnerships? Check. Strategic agencies were onboard.

  • Backing? Check. Government agencies invited them for soft launches; investors were circling.

So, why kill it?

The Trap Momentum is seductive. When things are working—when people are patting you on the back and offering you checks—it feels impossible to stop. You feel ungrateful. You feel crazy.

But Vivian realized something crucial: This project didn't align with the long-term leverage she wanted to build. The operational synergy was there, but the mission alignment wasn't. She looked at the future this startup would create for her life, and she didn't feel excitement. She felt trapped.

The Courage to Quit Stopping a failing project is easy; the market kills it for you. Stopping a successful project requires an immense amount of internal clarity.

Cermini AI is dead, not because it failed, but because its founders had the courage to say "No" to a good opportunity so they could keep their hands free for a great one.

💼 LinkedIn Post

"The Momentum Trap."

Vivian shared a hard truth this week. Her project, Cermini AI, had everything a founder wants: ✅ Strategic Agency Partners ✅ Government backing ✅ Funding interest

And she shut it down.

Why? Because it wasn't creating the right long-term leverage.

It’s easy to quit when things are failing. It is incredibly hard to quit when things are working, but working towards the wrong future.

Don't let momentum decide your direction.


3. Jun: AI CRM for WhatsApp

🔍 Overview

  • Project: Personal CRM extracting "Jobs to be Done" from WhatsApp.

  • Goal: Nurture connections by surfacing buried tasks/requests.

🎬 Reels Script

Hook: "WhatsApp is where business happens. It's also where leads go to die." Shotlist:

  • [0:00] Fast scroll through a messy WhatsApp inbox (blur names).

  • [0:03] Text overlay: "I forgot to reply... again."

  • [0:05] Screen recording: Jun's tool scanning contacts.

  • [0:08] A "Job List" pops up: "Book facility for Mark," "Reply to Sarah."

  • [0:11] One tap "Draft Reply."

  • [0:14] Cut to face: "Turn your chaos into a CRM."

Caption: Stop ghosting your network. 👻➡️🤝 #Productivity #WhatsApp #CRM #Networking

📝 Blog Post

Title: The Hidden "Jobs to be Done" in Your WhatsApp

If you live in Southeast Asia, WhatsApp isn't just a chat app. It's your operating system. It's where deals are closed, where friendships are maintained, and where favors are asked.

But as a workflow tool, it’s a disaster.

The Psychology of Message Debt We've all been there: You open a message from a client or friend. They ask for something simple—"Can you check on that booking?" or "Let's grab coffee next week." You’re busy, so you think, I’ll reply later. Then, 50 other messages come in. The request gets buried. The "later" never comes. This is "Message Debt," and it eats away at your professional relationships.

Turning Chat into Action Jun is building a solution that treats WhatsApp less like a stream of consciousness and more like a database of intent. His tool sits on top of your chats and uses AI to extract the "Jobs to be Done."

It scans the chaos and presents you with a clean list:

  • Mark needs the facility booking code.

  • Sarah is waiting for the pitch deck.

It turns the passive act of "checking messages" into the active workflow of "clearing tasks." In a world where responsiveness is often equated with competence, this tool is a superpower for anyone in sales or partnerships.

💼 LinkedIn Post

Networking isn't about collecting contacts. It's about maintenance.

But maintaining 500+ relationships on WhatsApp is a nightmare. Messages get buried, promises to "check on that" get forgotten.

Jun is building a layer that sits on top of WhatsApp to solve this. It scans your chats and extracts a "Jobs to be Done" list.

• Friend asked for a booking? -> It's on the list. • Client asked for a follow-up? -> It's on the list.

Simple. But for anyone in sales or partnerships, this is a superpower.


4. Andre: House Lens (Automating Media)

🔍 Overview

  • Project: House Lens (AI Newsroom).

  • Problem: Digital publishing requires huge teams (20-50 ppl) for manual workflows.

  • Solution: Automate Research -> Validation -> Production based on "Brand Voice."

🎬 Reels Script

Hook: "This AI just replaced a 20-person newsroom." Shotlist:

  • [0:00] Split screen: Left side "Old Way" (stock footage of busy office), Right side "New Way" (Andre's dashboard).

  • [0:04] Screen recording: Selecting a trending story.

  • [0:06] AI analyzing "Social Velocity."

  • [0:08] One click: Generates video script + article in the specific "Brand Voice."

  • [0:12] Text overlay: "From Story to Content in 30s."

Caption: The future of media isn't more people. It's smarter workflows. 📰🤖 #MediaTech #Automation #AI

📝 Blog Post

Title: The One-Person Newsroom: Automating Editorial Intuition

The traditional media model is built on friction. A story breaks. A researcher finds it. An editor validates it. A writer drafts it. A sub-editor polishes it. A producer films it. By the time the content is live, the trend is dying.

This week, Andre showed us "House Lens," and it felt like looking at the end of that old assembly line.

Codifying "Voice" The biggest challenge in automating media isn't generating text; it's generating personality. Generic AI content reads like... generic AI content. House Lens solves this by analyzing a publisher's historical archive. It deconstructs their "Voice Profile"—the slang they use, the sentence structure they prefer, the specific way they hook an audience.

Velocity at Scale The tool scans for stories with high "Social Velocity," validates them for accuracy, and then instantly spins up content that sounds exactly like the publisher.

  • For a Gen Z news site? It uses memes and quick cuts.

  • For a financial journal? It uses data and formal analysis.

We are moving away from the era of "AI as a Writer" and entering the era of "AI as Editor-in-Chief." This doesn't just cut costs; it allows a single creator to operate with the output of a 20-person newsroom.

💼 LinkedIn Post

The traditional newsroom model is broken. Research -> Validation -> Editorial -> Production.

That workflow usually takes a team of 20. Andre is building a platform that does it with 1.

Meet "House Lens." It doesn't just write articles. It analyzes a publisher's past content to build a "Voice Profile"—understanding their tone, slang, and audience.

It spots a trending story (based on social velocity), validates it, and generates the content in that specific publisher's style.

We're moving from "AI as a writer" to "AI as an Editor-in-Chief."


5. Faw: Cleve Engineering Update

🔍 Overview

  • Project: Cleve.ai (The product itself!).

  • Update: Major refactor to Convex, Zero Data Retention (ZDR), and Agentic Email.

🎬 Reels Script

Hook: "Why we rewrote our entire codebase." Shotlist:

  • [0:00] Faw looking tired/stressed (acting): "Tech debt. It's a killer."

  • [0:03] Transition: Snap fingers to new UI.

  • [0:05] Screen recording: Typing on one device, instantly appearing on another.

  • [0:08] Text overlay: "Real-time Sync (Convex)."

  • [0:10] Cut to "Privacy Mode" toggle.

  • [0:12] Voiceover: "And we stopped storing your data. Zero Data Retention."

  • [0:15] Face to cam: "Faster. Safer. Smarter."

Caption: Sometimes you have to tear it down to build it right. Cleve V2 is here. 🚀 #Engineering #Refactor #PrivacyFirst

📝 Blog Post

Title: Zero Data Retention: Why Privacy is Our New North Star

In the rush to build "smarter" AI, most companies are hoarding your data. They store your prompts, index your notes, and train their models on your private thoughts. We decided to go the other way.

This week, we rolled out a complete rewrite of the Cleve infrastructure, and while the speed improvements (thanks to Convex) are great, the real story is what isn't there: Your data.

The "Zero Data Retention" (ZDR) Philosophy We have moved our core processing to Base 10. This allows us to implement a strict ZDR policy.

  1. You send a request.

  2. We process it.

  3. We forget it.

We do not use your thoughts to train our models. We do not store your raw prompts in a way that can be mined later.

Why Convex? We also tackled the "refresh button" problem. By moving to Convex, Cleve now feels alive. If you update a note on your phone, it changes on your laptop instantly. No spinning wheels, no sync conflicts.

If we want AI to truly function as a "second brain," it needs two things: It needs to be as fast as thought, and it needs to be as private as your own mind.

💼 LinkedIn Post

We just rewrote Cleve from scratch. Here is why: Privacy.

Most AI wrappers are leaking data like a sieve. They store your prompts, your notes, your thoughts.

We decided to go the hard way: Zero Data Retention (ZDR). We moved our infrastructure to Base 10. When you use Cleve, we process the request, and then we forget it. We don't train on your data. We don't store your raw prompts.

We also moved to Convex for real-time syncing. No more refresh buttons.

If we want AI to be a true "second brain," it needs to be as private as your first one.