Field CTO, APAC @ Autohand AI | Ex-GitHub, Apple. Busy. Happy. Built a house. Purple. 💜😈👾
Leura, Australia
Joined May 2009
- Tweets3.9K
- Following529
- Followers253
- Likes381
We're a Launchpad Partner for AI Engineer Sydney 🚀
7–8 Dec, Hilton Sydney. Coding agents, evals, context & memory, and shipping AI that holds up in production.
Early bird ends 20 Nov. Come say hi 👋
webdirections.org/ai-enginee…
#AIEngineer
Part four of my Routes series for @autohandai: scoring, evals and trying it with your team.
How a route is scored, how to tune it without touching code, and how to measure whether it got better.
autohand.ai/updates/routes-p…
New to Routes? The whole series:
Part one: autohand.ai/updates/routes-p…
Part two: autohand.ai/updates/routes-p…
Part three: autohand.ai/updates/routes-p…
Part four also looks at decision models, including our own Weka, as a possible classifier. That's a direction we're exploring, not a promise.
Questions and ideas: discord.gg/ZM3TCtwCwG
The honest bit: Routes' built-in classifier reads the whole agent conversation, so nearly every agent request looks hard. It's the biggest open problem in Routes.
The most useful thing you could send us is real agent requests, labelled:
github.com/autohandai/routes…
![[Feature] Classify the latest request, not the whole agent conversation · Issue #9 · autohandai/r...](https://pbs.twimg.com/card_img/2118592651788156928/gTUoWphs?format=jpg&name=800x419)
Part three of my Routes series for @autohandai: one ask, a squad of agents.
Squad decides who does the work. Routes decides which model does the thinking. Give each role an outcome, not a model.
autohand.ai/updates/routes-p…
New to Routes? Start with part one, then part two on policies and guardrails:
autohand.ai/updates/routes-p…
autohand.ai/updates/routes-p…
Know that a Squad lead hands work out like a general contractor, but the plan lives in its reply. Nothing yet tracks the parts of one job across the team.
That's the gap we'd most like your ideas on - Squad issues.
Routes: github.com/autohandai/routes…
Squad: github.com/autohandai/squad/…
Part two of my Routes series for @autohandai: what changes with more than one model to choose from. Policies, gates, failover, guardrails.
The honest bit: router-privacy is a strong preference, not a wall. Here's how to keep work local for certain.
autohand.ai/updates/routes-p…
New to Routes? Part one gets it running on your machine, from install to pointing an agent at it:
autohand.ai/updates/routes-p…
Routing is a busy field, so part two also puts Routes next to OpenRouter's new Jev Router: what each does better today, and how to run a hosted router as one of the models behind Routes.
Questions welcome: discord.gg/ZM3TCtwCwG
Routes is in very early beta, and part two ends with open questions:
Should auto follow the config's policy line?
Should there be a strict local-only mode?
Would budget windows help?
Which coding prompts does redaction mangle?
Each has an issue: github.com/autohandai/routes…
Daniel Figucio retweeted
I’m not a visual learner. One of my superpowers is understanding what you’re going to say before you finish saying it.
But I know plenty of people who learn better when they can see and touch things.
Today, I’m making Fluid UI available to anyone with an account on @autohandai.
dev.autohand.ai
Fluid UI turns AI responses into interactive experiences directly inside chat. Instead of just reading text, you can explore visual interfaces, change inputs, interact with components, and see results update in real time, all within the conversation.
The goal is simple: make AI responses something you can interact with, not just read.
The funny part? Fluid UI has been there since August. The world just hadn’t discovered it yet.
I built this example while waiting for my baby to wake up. I’m sitting in the car, making things.
That’s building a company. Sometimes, all you need is an idea and a little time.
Go there and enable Fluid ui in experiments.
That’s my trying to understand how a plane can have midair control on velocity and breaks.
autohand.link/s/ah-4agv2h36m…
Routes puts one OpenAI-happy URL in front of all your models. Ask for an outcome (balanced, cheapest, fastest, strongest) and it picks a model for each request, and tell you why.
Very early beta. Part one gets it going. Tell us what breaks.
autohand.ai/updates/routes-p…
@autohandai
Complex engineering work needs room to unfold. Meet Autohand Code Pro and Code Max - new plans that give developers more capacity to stay in flow, work through complexity and carry substantial tasks through to completion!
autohand.ai/updates/introduc…
@autohandai
@AssistDelta who can I get to contact me about a luggage that ended up in your LAX Baggage service office because it was put in baggage claim at LAX instead of being transferred to my flight to SYD even though it was checked through? I'd like it on the next QF flight to SYD...
Proud to be the 58th 😎 on @BackerKit for Dungeon Murder: A New Game By Scott Johnson. dungeon-murder-a-new-game-by…
Proud to be the 1,703rd 😈 on @BackerKit for Terraforming Mars: The Dice Game. Thanks @strongholdgames terraforming-mars-dice-game.…
Daniel Figucio retweeted
GitHub Issues – Projects now generally available github.blog/changelog/2022-0…
