Sign in to view Andris’ full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Sign in to view Andris’ full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Toronto, Ontario, Canada
Sign in to view Andris’ full profile
Andris can introduce you to 10+ people at Better Stack
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
1K followers
500+ connections
Sign in to view Andris’ full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
View mutual connections with Andris
Andris can introduce you to 10+ people at Better Stack
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
View mutual connections with Andris
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Sign in to view Andris’ full profile
or
New to LinkedIn? Join now
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
About
Welcome back
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
New to LinkedIn? Join now
Activity
1K followers
-
Andris Gauracs shared thisI got a 45 million parameter AI model running fully offline on a $10 microcontroller. The model is Needle 2, an open source release from Cactus Compute. It's 14MB, Apache 2.0 licensed, and built specifically for tool calling rather than general chat. In theory it's small enough for edge hardware, but Cactus doesn't officially support bare microcontrollers like the ESP32-S3. So I figured I'd try building my own inference engine to get it running there anyway. And it actually worked. I can type something like "flash a red light for three seconds," and watch the model reason through it, generate a valid tool call, and execute it, all on the device itself. No WiFi, no cloud, nothing. It even gives you a live confidence score for every answer, and when I asked it something completely out of scope, like the capital of France, it correctly figured out it had no tool for that instead of just making something up. Honestly, this project cements my belief about where on-device AI seems to be headed. It's not about bigger models that need more and more compute, but rather smaller, narrower ones that are really good at one specific job and perform them with near perfect accuracy. I imagine this could be a really good baseline for building DIY robotics projects. I've added the link to the GitHub repo in a comment below.
-
Andris Gauracs reposted thisAndris Gauracs reposted thisWe ran a 45M parameter AI model fully offline on a $10 microcontroller. No WiFi, no cloud. It reasoned through natural-language commands and gave live confidence scores. Watch the full breakdown.
-
Andris Gauracs reposted thisAndris Gauracs reposted thisDecimen sends files using nothing but a screen flashing QR codes and a camera reading them back. No WiFi, no Bluetooth, no network at all. We tested it and got 3 KB/s instead of the claimed 128 KB/s. We dug into the source code to find out why.
-
Andris Gauracs reposted thisAndris Gauracs reposted thisThinkingCap cuts AI reasoning tokens by up to 58% with basically no accuracy loss. We break down how BottleCap trained it and ran it ourselves against the base Qwen3.6-27B model to see how good it is.
-
Andris Gauracs shared thishandy has been my absolute favourite tool lately. 👋Andris Gauracs shared thisWe put a free, open-source dictation app up against Google voice typing and Whisper Flow using the same test passage. It held its own against the paid tool and beat Google by a mile. Offline, no subscription, no account. Full breakdown in the new video.
-
Andris Gauracs reposted thisAndris Gauracs reposted thisWe ran a 28.9M parameter LLM on an $8 chip with 512KB of RAM. Broke down the memory trick that makes it possible, rebuilt the whole pipeline from scratch, and tested it with a few custom prompts of our own.
-
Andris Gauracs reposted thisAndris Gauracs reposted thisJack Dorsey's Block just dropped Buzz, an open-source Slack and GitHub rival in which your AI agents are full members with their own cryptographic identities. We took it for a spin. Here's what we found.
-
Andris Gauracs reposted thisAndris Gauracs reposted thisWe ran a 744-billion-parameter AI model on a laptop using a tiny C tool called Colibrì. Testing it on a MacBook vs an RTX 5090 revealed the real bottleneck for local AI, and it isn't the GPU.
-
Andris Gauracs reposted thisAndris Gauracs reposted thisWe think we just found the best file transfer tool out there. It's called croc, it's free, open-source, and beats AirDrop, WeTransfer, and scp at their own game. Full breakdown and demo in our latest video.
-
Andris Gauracs reacted on thisAndris Gauracs reacted on thisStill playing around with Claude and finding new ways to put it to work 😅 This time: Claude Code + Blender. I went for a walk, and Claude stayed home making a 3D model for me. It would send me screenshots, I’d leave feedback on shapes and details, and it kept refining the model. By the time I got home, the model was done and animated. The only things I touched manually were the materials and lighting, because the first render had a bit of a play-doh situation going on 😁 I also gave myself a hard limit of ~1.5 hours for the whole experiment, so the cable still feels a bit rigid and isn’t perfect, but overall I’m pretty happy with it. Video below 👇 #AIDesign #ClaudeAI #Blender
-
Andris Gauracs reacted on thisAndris Gauracs reacted on thisAt this point, Claude needs a translator for Claude. As a non-native English speaker, I genuinely thought my brain was just tired every time Opus 5 gave me a paragraph that needed three rereads to decode. But the slightly scary part is this: when you spend hours reading and writing with an LLM, its style starts rubbing off on you too. We trained AI on human language. Now AI might slowly start training human language back. And apparently, the first dialect is Claudish.
-
Andris Gauracs reacted on thisAndris Gauracs reacted on this🐶 WE ACTUALLY DID IT. We built a Magical Dog Bus. 🚌 Meet the real security pack behind the magic at Wiz. Every morning, they hop on the bus, roll into work, and protect everything you run, build & fetch. 🐾 Yes, they really ride the bus & honestly, they're overqualified. Happy Dog Day from Wiz and our very best threat sniffers. 💙
-
Andris Gauracs reacted on thisAndris Gauracs reacted on thisHow to make an atomic bomb? ☢️ Local models like Qwen 3.8 27B can be modified to remove any kind of guardrails. This is achieved by suppressing the model censorship components with tools like Heretic: https://lnkd.in/ehR9KwnC I asked how to make an atomic bomb and other stuff, It gave me very cool tutorials 😂 Open weight models, which are amazing, are also a double-edged sword.
-
Andris Gauracs liked thisAndris Gauracs liked this𝗡𝗼𝘁 𝗝𝘂𝘀𝘁 𝗔𝗻𝗼𝘁𝗵𝗲𝗿 𝗦𝗼𝗹𝗶𝗱𝗮𝗿𝗶𝘁𝘆 𝗦𝗲𝗹𝗳𝗶𝗲 Zelenskyy with the leaders of Denmark, Estonia, Finland, Latvia, Lithuania and Norway in Kyiv today, but the photograph captures something more important than another demonstration of political support. Ukraine is no longer only receiving European security; it is increasingly helping shape it, as Europe provides weapons, financing and industrial scale while Ukraine provides battlefield experience, rapid adaptation and lessons European militaries are now actively trying to absorb. 𝘛𝘩𝘪𝘴 𝘪𝘴 𝘯𝘰 𝘭𝘰𝘯𝘨𝘦𝘳 𝘫𝘶𝘴𝘵 𝘢 𝘤𝘰𝘢𝘭𝘪𝘵𝘪𝘰𝘯 𝘴𝘶𝘱𝘱𝘰𝘳𝘵𝘪𝘯𝘨 𝘜𝘬𝘳𝘢𝘪𝘯𝘦; 𝘜𝘬𝘳𝘢𝘪𝘯𝘦 𝘪𝘴 𝘣𝘦𝘤𝘰𝘮𝘪𝘯𝘨 𝘱𝘢𝘳𝘵 𝘰𝘧 𝘌𝘶𝘳𝘰𝘱𝘦’𝘴 𝘥𝘦𝘧𝘦𝘯𝘤𝘦 𝘢𝘳𝘤𝘩𝘪𝘵𝘦𝘤𝘵𝘶𝘳𝘦.
-
Andris Gauracs reacted on thisAndris Gauracs reacted on this🚨 University of California, Berkeley JUST OPEN-SOURCED FREETOKEN, AND THE RESULTS ARE WILD It is a new local inference engine running 2-4x faster than @ollama by exploiting Mixture-of-Experts architectures. The initial benchmarks point to a massive shift for local capabilities: → Qwen3.6-35B on an 8GB GPU at 39.3 tokens/s → DeepSeek-V4-Flash 284B on a 32GB GPU at 22 tokens/s → GLM-5.2 753B on a 96GB GPU at 14.9 tokens/s A 35B model normally requires 70GB for weights. FreeToken serves it on an 8GB GPU because compute is no longer the bottleneck. Instead of loading the full model: > it targets the MoE router. > it measures your machine's bandwidth to dynamically split memory misses between the CPU and PCIe. This architecture is also a massive win for agents. Coding agents constantly rewrite history, forcing thousands of tokens through prefill. FreeToken saves checkpoints exactly at agent framework boundaries, dropping first-token latency from 232 seconds (llama.cpp) to under 44 seconds. Apache 2.0. OpenAI and Anthropic API compatible. 100% free and open-source. Repo: https://lnkd.in/ec88AU3E Paper: https://lnkd.in/ey9PURfd Shoutout to University of California, Berkeley for building this and making it open-source for the community 🤗 Don't forget to drop a ★!
-
Andris Gauracs reacted on thisIf your girl: - drinks a lot of water - causes drama - remembers everything That's not your girl, that's a data center
-
Andris Gauracs reacted on thisAndris Gauracs reacted on thisbrowser-use finds the button on the page instead of memorizing a selector, so a renamed div doesn't turn your test suite red. On Odysseys, a benchmark of 200 long-horizon, multi-site web tasks, it ranks second overall and beats every computer-use agent OpenAI, Anthropic, and Google built for their own frontier models. Page object models were always a coping mechanism for agents that couldn't see the page. 110k stars, MIT. #AIAgents #BrowserAutomation #TestAutomation #OpenSource
Experience & Education
-
Better Stack
********* ********
-
**** *******
******** *******
-
******
****** ******** ********
-
***** ******** ************
******** ****** ******** ****** undefined
-
-
***** ******** ************
********** ****** ******** *** *********** ********* *******
-
View Andris’s full experience
See their title, tenure and more.
Welcome back
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
New to LinkedIn? Join now
or
By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.
Honors & Awards
-
Nomination: Best editing director
Lielais Kristaps 2014
Nomination for the "Best editing director" award for the documentary film „Fedja” - Andris Gauračs & Kristīne Želve
-
1st place at "Tieto24 - 24 hour business idea contest"
Tieto Latvia
Award was won by a team who developed the best social mobile app idea. I was the captain of the team "Oranges" and our idea for a new eco-friendly social network app won the award.
Languages
-
English
Full professional proficiency
-
Latvian
Native or bilingual proficiency
-
French
Limited working proficiency
-
Russian
Elementary proficiency
Recommendations received
1 person has recommended Andris
Join now to viewView Andris’ full profile
-
See who you know in common
-
Get introduced
-
Contact Andris directly
Explore more posts
-
Toya
2K followers
For mobile game developers, moving to Roblox shouldn’t mean giving up the infrastructure and visibility they rely on. Roblox’s native toolset wasn’t designed to fit seamlessly into the professional studio workflows established by mobile game publishers. PlayTiger bridges that gap with an integrated CI/CD pipeline, full-funnel attribution, and measurement analytics built for Roblox - giving teams familiar control over deployment, acquisition, player behavior, and campaign performance. The result: Roblox becomes a measurable, scalable publishing channel - not a black box.
Explore collaborative articles
We’re unlocking community knowledge in a new way. Experts add insights directly into each article, started with the help of AI.
Explore More