Does this include a turn taking API? It'd be great to have one API that could do "Conversation in a box". One of the biggest annoyances is daisy chaining many models together for turn taking, dumb models for immediate responses, with smarter models returning and taking over after.
Yes, on the hosted side (agents platform): full sessions come with VAD and turn-taking handled - we set them up and tune them for your use case, so that is the closest thing to conversation in a box. If you run your own orchestration, the gateway is just the routing layer and turn-taking stays in your framework - in our own stack we run Pipecat's Smart Turn in-process and tune the commit threshold on real calls. We also share our benchmarks here: https://benchmarks.speko.ai/turntaking
On fast dumb models answering while a smarter one takes over: we are experimenting with exactly that split - a small fast model holds the conversation while a larger one works behind it. Today it runs as two pinned routes, not one packaged API. Most turns in a phone call do not need a frontier model, and the fastest models on our LLM board are all small, so this is where routing earns its keep. We publish benchmarks on LLMs here: https://benchmarks.speko.ai/llm
One packaged API for the takeover would be great, also worth considering predicting answers with a stronger LLM while the user is talking. Ello (the reading tutoring company) does this.
Built in AI editor for Wordpress to make it easy to make your site better. Originally just for online newspapers but starting to get interest from Wordpress agencies. First user is free: https://presspass.ai/ai-editor
Highly rec going after a specific vertical - healthcare might be the right spot given your experience. Why did you use DuckDB instead of CockroachDB/Snowflake?
Our outreach is vertical-specific, and healthcare is indeed on the list! But what we learned working a vertical is that the primitives underneath (shared queries, permissions, caching, refresh semantics) repeat across industries.
We use DuckDB internally because we like its ergonomics - it's flexible, runs well in memory, manages a lot of file structures under the hood, but we do work with Snowflake (and Databricks and other warehouses) as well.
If you think this is stupid or you know of a more annoying “click based” workflow that should be automated, let me know! I’m early and need more thoughts.
One thing I remember from my days in the LMS world is that obfuscated copies of prod tenants were used for testing. Almost every dev had at least one tenant from prod on their local computer. So with some de-obfuscation at least some of the data is plausibly retrievable. Whether that data is also public depends on how the negotiations go.
The death of local journalism is fundamentally a revenue problem. My cofounder and I have been working for the last year to find new revenue streams for newspapers at https://seward.presspass.ai/.
Our current hypothesis is that local rewards programs could be a sustainable revenue stream and give the newspaper a way to prove their advertising works with locals.
While trying this out, we've also helped a few papers get up and running - we're calling it "newspaper in a box". Check out a few of the papers we've helped launch: https://sewardfolly.com/ (9 months old) https://homerindependentpress.com/ (2 weeks old).
Here's a couple concrete examples I found. It seems like most have to rely on multiple revenue streams for it to work -- something like 1/3 reader revenue (membership), 1/3 grants/philanthropy, 1/3 earned income (events, merchandise, etc).
Small team does local journalism, motivated locals donate to keep it running. It's that simple. Some people are happy to donate $20 a month to their favorite Twitch streamer or open source project, and other people are happy to do the same for their local newspaper.
The failed model is trying to run it like a journalism factory: producing articles at some marginal cost and selling them at a fixed price that exceeds marginal cost.
Just look at NPR and member stations. The federal government ended their funding, but they kept right on going because of donations.
From here you can see there's various references and entry points to hyperlocal journalism in the US, Canada, etc, all working to fill the void of local newspapers departing, or local newspapers increasingly only providing national or state level news.
Instead of ads alone, its a sponsorship model, one from the readers side (optional), and one from the local organizations who are outright sponsoring the local news to be created, maybe in different categories, etc.
I'm working on creating Lovable for popular CMSs like wordpress, ghost, substack etc.
I've spent that last year working with news organizations to improve their online presence and a lot of my job has ended up being translating their ideas into wordpress clicks using Claude or ChatGPT. They can't take advantage of products like Lovable because the resulting code won't work with the CMS that their staff is used to.
Email me if you'd like free access! seamus@presspass.ai
I still want something no one has, which is the ability to launch agents in different git worktrees simultaneously and check the results out on my main branch for testing when they are finished.
does minimal overhead with agent orchestration (its just a bash/typescript) as its main focus was adding enhancements to codex like double redundant checkpoint via git and jj (lessons learned from codex being git reset --hard happy), something like claude skills (just a bunch of mds that steer it towards specific activity like think, plan, execute), timeout wrappers (to get you unstuck if codex waits a long time), blacklist commands during yolo (rm -rf, git reset banned even if it by small chance run it) MIT licensed
you can work sequentially (subagents launch one after the other) or parallel (worktrees) but tbh sequentially is better because you understand what is going on with parallel it might be best for dealing with tests and UI.
I am curious: why would you you like to have that? (Genuine question, I am personally so scared about the AI going crazy and putting slop everywhere that I often ask it to focus on a single well defined area first)
