Erick

AI Engineer | Custom agents & automations | Sharing live builds

[Ra @slimepriestess]

me and my friends have taken a somewhat different trajectory

[Garry Tan @garrytan]

You can basically work with agents to write markdown skills and put it on a cron job and do almost every useful type of knowledge work now

[Erick @aibyerick]

@infiloop2 agreed. three things make it survivable in dograh: tools are scoped per step, a failed tool result is flow input you route (retry, human transfer, clean hangup) instead of something the agent improvises around, and every turn streams live so you hear it as the caller does.

[Erick @aibyerick]

day 9 shipped dograh-ctl to production

- dograh is an open-source, self-hosted voice agent platform. it has an api, sdks, and a dashboard, but no command line. so i built one: 32 commands, agents, numbers, models, campaigns, call runs, all from the terminal.
- claude code can now create, publish, number, test, and call a dograh voice agent end to end, and a phone agent can be driven mid-call.

[Erick @aibyerick]

@Joshua_WD Agree! composable was the goal: every command takes --json, the dangerous ones ask before they cost money, and the whole thing runs as an mcp server so an agent can drive the platform the same way a human does from the terminal

[Erick @aibyerick]

thx! I’ve been building voice agents for the last 2 years. mid-call is handled by the agent's flowchart: every step has its own instructions and a set of tools it can call while talking (hit an api, call an mcp server, transfer, hang up). the tool result steers which step comes next, so the agent acts in the middle of the conversation instead of just chatting. the cli wraps the rest. happy to trade notes on the multi-channel build

[Erick @aibyerick]

@SEOMastery2026 Thx to Fable 5!

[Erick @aibyerick]

open sourced it here: https://t.co/WDplYoiXja. pip install dograh-ctl

[Lukas Ziegler @lukas_m_ziegler]

🚨 BREAKING:

@GravisRobotics has raised $200M in the largest Series A in construction robotics history, led by SoftBank! 🔥

Spun out of ETH Zürich in 2022. Three years later, $200M and SoftBank backing.

The company is bringing autonomous AI to the most physically demanding machines on earth. Excavators. Heavy construction machinery. Equipment that hasn't fundamentally changed in 50 years. 🚜

Most physical AI operates in static worlds. Self-driving cars navigate around obstacles. Warehouse robots move objects across fixed floors. Everything stays where it is.

An excavator does the exact opposite.

It works by intentionally crashing into the environment, breaking apart soil with hidden rocks, reshaping the earth with every single pass of the bucket.

The world changes with every action the machine takes.

→ Gravis AI doesn't navigate a static world, it actively takes the world apart and puts it back together
→ Models trained on billions of cubic yards of simulated earth, from soft clay to rock-filled soil
→ Generalises across different machine manufacturers — not locked to a single platform
→ Brings factory-floor precision to historically unpredictable civil jobsites

The macro case is overwhelming. Energy networks, data centres, housing, transit, climate infrastructure, all of it requires construction at a scale the existing workforce cannot deliver.

Construction is the primary bottleneck of the entire physical AI economy.

The same AI boom driving demand for data centres is now funding the robots that will build them.

Europe keeps producing world-class deep tech. 🇨🇭🇪🇺

To the team behind this, MASSIVE CONGRATS!

Can't wait to publish what we have created onsite with Gravis team! 🫶🏼

~~

♻ Join the weekly robotics newsletter, and never miss any news → https://t.co/GoA3ZuwoPB

[🔨Robert The Builder 🇺🇸 @NobodymrRobert]

It's happening.

[stash @stash_pomichter]

Announcing the first production robot navigation framework on $500 hardware

Explore the world once → your robot agent will relocalize and build a persistant, spatial memory across sessions

SLAM, relocalization, loop closure, map i/o, planning, control

No ROS. Open source.

[Erick @aibyerick]

@zanehengsperger The Claude code effect 👀

[Erick @aibyerick]

day 8: i built a command line for dograh so i can split test voice ai stacks in one place

- every real-time voice platform ships its own cutting edge models, and comparing them head to head meant hand building api infra against each one
- voice is going vertically integrated: the llm, the transcriber, and the voice model becoming one stack instead of three separate pieces you wire up yourself
- dograh is the open source platform that can run those stacks side by side, it just had no command line
- so i built one. now my agents orchestrate the integration and split test new models against my current setup, from the terminal

[Erick @aibyerick]

day 6 i built a pipeline powered by Opus 4.8 (until we get fable 5 back) that generates and films websites, then has Gemini Pro 3.1 grade the footage

- a remotion pipeline programmatically shoots the live site into a cinematic product reel. code renders the video, not a screen recorder.
- then gemini 3.1 pro watches the reel and hears the audio, and scores it against a our evals. it sees what opus 4.8 can't.
- the agents build the site, film it, and judge the film.

[Erick @aibyerick]

day 5 - i wired claude fable 5 and gemini 3.1 pro into one self-grading pipeline

- fable 5 built a world-class landing page in three.js, real motion, live in the hour.
- but fable 5 isn't multimodal, it literally can't see what it ships. so it was building blind.
- so gemini 3.1 pro became its eyes: it records the live page, scores it against a world-class bar, hands back the one fix that matters, and loops until it passes.
- two rival labs' models as one line. i open sourced the whole gate. the agents QC themselves now.

[Erick @aibyerick]

@JustinGorya @DevinAI Agree, here is how it did with music, animation, graphics *chefs kiss*

https://t.co/gd9L6qrf3f

[Erick @aibyerick]

Fable 5 is honestly surreal.

I asked it to generate the most optimal focus music.

This was the output...

[Erick @aibyerick]

Fable 5 is honestly surreal.

I asked it to generate the most optimal focus music.

This was the output...

[Erick @aibyerick]

day 4 testing fable 5 on my music generation pipeline
- my first music mix got pushed hard out of the gate by youtube, a couple thousand impressions. the demand is real, the only question is what to make.
- so i built an agent to find out: it searched github, integrated an open-source youtube tool, and ran data pipelines to validate the niche with real search data.
- it upgraded the look and feel against real design principles, generated, not guessed.
- the agents now QC themselves. every asset one model makes, another model judges

[Erick @aibyerick]

day 4 testing fable 5 on my music generation pipeline
- my first music mix got pushed hard out of the gate by youtube, a couple thousand impressions. the demand is real, the only question is what to make.
- so i built an agent to find out: it searched github, integrated an open-source youtube tool, and ran data pipelines to validate the niche with real search data.
- it upgraded the look and feel against real design principles, generated, not guessed.
- the agents now QC themselves. every asset one model makes, another model judges