Get live statistics and analysis of haugas.dev's profile on X / Twitter

An idea that's been relentlessly bouncing around my head lately: the MCPUI concept I recently posted about. SSE can carry detailed hints about parameter types. With this information, rendering the UI becomes relatively easy, making interaction with the LLM more engaging and enjoyable. With even this minimal information, the LLM can dynamically render a full UI on the fly based on user preferences and possibly some background documentation of available UI systems. This might involve a side agent specialised specifically in quickly generating the right UI, possibly including a caching mechanism. That's the MCP SaaS, where each user can simply select which interface they want to use, e.g., a chat-based interface with partially rendered tool actions or maybe a simple prompt: "render me the admin panel of service X, check my preferences...". So many possibilities here: who likes brown, who likes pink, who prefers a touch interface, or perhaps another user likes quick voice commands. But in the end, regardless of the input, the system understands you.

Cursor went down today, and I learned something. I tried to refactor some code manually while it was offline. Almost half of my AI-generated code was useless. I have become so dependent on AI that I have ignored deep reviews, and it's rather rare. But... here's the thing: manual review only takes a couple of minutes and catches so much. It's a good reminder for me to do deep code reviews more often.

VoltAgent = IKEA instructions for AI agents…

👀

I showed up. I also took part in the Tallinn Cursor meetup's mini hackathon. I won. I consistently burned tokens so much that I became the top 1% of the cursor token burners in my area. But I also used my skills that I consistently have improved 2 years. So what did I learn? Keep showing up, keep burning the tokens, keep winning, don't break the circuit.

Ever notice how AI assistants handle mistakes better than most humans? me: u did cause a lil fuckup though: [pasted error msg] assistant be like: "Ah shit, I caused a TypeScript error! The issue is on line 61 where I'm trying to call Object.keys() on parsedContext, which could be undefined." I love how AI admits mistakes like a junior dev who sincerely cares 😂

MCPs are trending. I love the concept of MCP-based SaaS because it's flexible and won't restrict the end user too much to a specific UI. You can plug it into any interface: voice, text-based, web-based, and so on. So the SaaS becomes less dependent on the UI.

$3k taught me everything about training AI for small languages. Here's how to do it for $1k. My client's telemarketing company wanted to push the boundaries. How far could we take AI phone calls? They wanted to explore something ambitious. AI capable of engaging in natural conversations in Estonian. Not just basic scripted responses, but real conversational AI that could adapt and respond naturally. So I decided to build the first AI that could fool native Estonian speakers in real conversations. I began with a 5-year-old proof-of-concept model developed at a local university. Completely different tech, just to validate the idea was possible. But for production, I needed to train from scratch using StyleTTS2. Two-stage process: Stage 1: Train the model to understand waveforms and generate clear, monotonous Estonian speech. Stage 2: Add prosody, naturalness, and context-aware tone adaptation. But first, I had to build the supporting pipeline. StyleTTS2 needs roughly 4 other components to work: 1. A speech recognizer to segment audio files. 2. A phoneme converter for Estonian sounds. 3. A prosody model for Estonian stress patterns. 4. A text corpus processor for training data. Each had to work perfectly with Estonian. Most existing models were built for English. I started the data preparation. The progress bar showed: 8 hours remaining. I immediately stopped and coded a parallel processor using all CPU cores. 8 hours became 15 minutes. Good thing I did. Because even with 15-minute data prep, I spent weeks of trial and error getting everything right. Different parameters, different preprocessing approaches, constant iteration. Imagine if each iteration took 8 hours instead of 15 minutes. I would still be waiting for my first successful training run. The lesson: When you see an 8-hour process, don't accept it. 15 minutes of coding saved me weeks of waiting. But the really expensive lessons were just beginning. Up next: Why 4xH100 GPUs produced garbage but 4xA100s worked perfectly.

vibe harder: good old Trello MCP for life 🔥

Teal is as close as you can get to seeing the new color without having your eyes lasered.

Indie vibes, tasty stack 👨🏼‍💻

1000 followers 🥳 Can't believe people like my weird tech ideas 😅 Thanks a lot!

VoltAgent = IKEA instructions for AI agents…

Claude 4 max mode reality check: $224 spent in 3 days with Cursor. Coding experience is amazing. Going back to regular mode is like using dial-up internet.

This is exactly why it still pays off to know how to code properly... The LLM did not get it wrong, all good! But it is unnecessary. The powerful concepts get lost in the dust. And it will compound quickly in a sufficiently large codebase.

Voice interfaces are having a moment. Just won a hackathon by building voice and MCP integration, while everyone else created dashboards. The pattern is clear. We're moving from clicking to talking. ElevenLabs is just the beginning.

MCPs are trending. I love the concept of MCP-based SaaS because it's flexible and won't restrict the end user too much to a specific UI. You can plug it into any interface: voice, text-based, web-based, and so on. So the SaaS becomes less dependent on the UI.

Indie vibes, tasty stack 👨🏼‍💻

Used DeeperSearch to find posts on X and got locked out for violating... whatever. Ironic how X own tools ambush users. Reactivation was easy though 🤷‍♂️

What I learned training Stage 2 of StyleTTS2. Last time, I told you about the H100 vs A100 GPU drama that cost me weeks of debugging. Finally achieved stable training on four A100 GPUs. Stage 1 was working. The AI could speak clear Estonian. But it still sounded monotonous and boring, nobody would bother to listen. Clear words, zero emotion, and no natural rhythm. Time for Stage 2 training. This adds prosody, emotion, and naturalness to the voice. Here's where things got slower. The Stage 2 training code couldn't use multiple GPUs. Had to switch back to single-GPU training. It took only a single day to train stage 1. Using only a single GPU meant waiting time tripled or more. Even during Stage 2, I discovered more data problems. Some of my text processing wasn't done correctly. The AI wasn't getting the right pauses because my punctuation processing was wrong. Commas, periods, and question marks - all crucial for natural speech rhythm. Started preprocessing the data again. Correct punctuation, proper sentence structure, clean context markers. The process was slow, but the model picked up the correct patterns again. Slowly, over many training iterations, something clicked. The AI started pausing naturally between thoughts. It emphasized the right words based on punctuation cues. What I learned: Stage 2 is all about the details. So the things I had to get right were punctuation, text formatting, and patience. Getting the data right matters more than speed. Total cost so far: Around $3k in GPU time and weeks of trial and error. But now I knew exactly what worked. So this is a short story on how. Perhaps I should create an agentic blueprint on how the process works, n8n for instance... Thinking about it.

An idea that's been relentlessly bouncing around my head lately: the MCPUI concept I recently posted about. SSE can carry detailed hints about parameter types. With this information, rendering the UI becomes relatively easy, making interaction with the LLM more engaging and enjoyable. With even this minimal information, the LLM can dynamically render a full UI on the fly based on user preferences and possibly some background documentation of available UI systems. This might involve a side agent specialised specifically in quickly generating the right UI, possibly including a caching mechanism. That's the MCP SaaS, where each user can simply select which interface they want to use, e.g., a chat-based interface with partially rendered tool actions or maybe a simple prompt: "render me the admin panel of service X, check my preferences...". So many possibilities here: who likes brown, who likes pink, who prefers a touch interface, or perhaps another user likes quick voice commands. But in the end, regardless of the input, the system understands you.

$3k taught me everything about training AI for small languages. Here's how to do it for $1k. My client's telemarketing company wanted to push the boundaries. How far could we take AI phone calls? They wanted to explore something ambitious. AI capable of engaging in natural conversations in Estonian. Not just basic scripted responses, but real conversational AI that could adapt and respond naturally. So I decided to build the first AI that could fool native Estonian speakers in real conversations. I began with a 5-year-old proof-of-concept model developed at a local university. Completely different tech, just to validate the idea was possible. But for production, I needed to train from scratch using StyleTTS2. Two-stage process: Stage 1: Train the model to understand waveforms and generate clear, monotonous Estonian speech. Stage 2: Add prosody, naturalness, and context-aware tone adaptation. But first, I had to build the supporting pipeline. StyleTTS2 needs roughly 4 other components to work: 1. A speech recognizer to segment audio files. 2. A phoneme converter for Estonian sounds. 3. A prosody model for Estonian stress patterns. 4. A text corpus processor for training data. Each had to work perfectly with Estonian. Most existing models were built for English. I started the data preparation. The progress bar showed: 8 hours remaining. I immediately stopped and coded a parallel processor using all CPU cores. 8 hours became 15 minutes. Good thing I did. Because even with 15-minute data prep, I spent weeks of trial and error getting everything right. Different parameters, different preprocessing approaches, constant iteration. Imagine if each iteration took 8 hours instead of 15 minutes. I would still be waiting for my first successful training run. The lesson: When you see an 8-hour process, don't accept it. 15 minutes of coding saved me weeks of waiting. But the really expensive lessons were just beginning. Up next: Why 4xH100 GPUs produced garbage but 4xA100s worked perfectly.

I built an AI that Estonians actually want to talk to. Estonian customers hate sales calls in English and hang up immediately. My client needed something completely different for their telemarketing company. I started with a free voice model from a local university. It was 5 years old and sounded monotonous, making it obvious you were talking to a bot. Not exactly what you'd call persuasive. But I had to start somewhere. I built the whole system around this basic voice first. Live transcription to understand customer speech. LLM responses are generated in real time. Background models to detect when customers say "yes" or "ok" so the AI doesn't interrupt them. Voice filtering and resampling. SIP phone integration for actual calls. Everything had to work perfectly, because even a single broken piece meant failed calls with real customers. Two years of constant testing, breaking, and fixing. Learning what Estonian customers wanted to hear and how they expected to be spoken to. The breakthrough came when I finally collected enough data to train my model. I used the StyleTTS2 architecture and spent one month training on 4 A100 GPUs. Many failed attempts with different parameters before getting it right. Today, my AI agent makes telemarketing calls in perfect Estonian. Customers can't tell it's AI, and they actually stay on the line to listen. Two years felt endless while I was building it. But now I have something nobody else has built for Estonian. Sometimes the hardest path teaches you the most.

I'm proudly the FIRST follower of upostify’s creator... Someone here definitely could use this new posting tool.

I showed up. I also took part in the Tallinn Cursor meetup's mini hackathon. I won. I consistently burned tokens so much that I became the top 1% of the cursor token burners in my area. But I also used my skills that I consistently have improved 2 years. So what did I learn? Keep showing up, keep burning the tokens, keep winning, don't break the circuit.

Ever notice how AI assistants handle mistakes better than most humans? me: u did cause a lil fuckup though: [pasted error msg] assistant be like: "Ah shit, I caused a TypeScript error! The issue is on line 61 where I'm trying to call Object.keys() on parsedContext, which could be undefined." I love how AI admits mistakes like a junior dev who sincerely cares 😂