Greetings, salutations, and all good things in between! (^_^)
For the past couple of years I’ve been prototyping and building out a solution for creating complex Ai workflow chains in my apps.
- We’re to a point now where we’re looking for case studies
- These early adopters will help us find holes we’ve missed, functionality that we should add, etc.
How it all came about
One of my long-term clients is a Dr. in New York, and one of our joint projects is an app that connects specialists with the care team for at risk patients in nursing homes. We track weights, meds, lab results, encounter notes… you name it, we track it; then each week we have specialists call in for a discussion with the care team at the facility.
When Ai came onto the scene, my client wanted to try and use Ai for everything - and I built this very complex ai processing chain system into the app.
- It was glorious
- All built inside AppSheet, all the complexities handled through automation
- This thing was a wonder of AppSheet engineering (if I do say so myself)
The problem was that it was all inside that ONE app; so when I wanted the same sort of ability in another app, I had to build it all AGAIN.
- And I hate replicating work
So I sought out a way to take everything from the nursing home app, all the workflow processing stuff, and build that into a separate app; this way I could do all the setup and configuration of the complex chains in the other app, and sorta… “weave” the two apps together.
Fast forward to today… now AppSheet still serves as the main UX, but everything is now powered through custom end points built out on GCP. My Dr. client has deeply integrated all of their systems with Aigents - forming a backbone of Ai automation.
- Analyze contents of emails
- Make records in the app
- Trigger further analyses
- Send emails
- Check if things are okay, or if something is off
- Generate summaries, extract data, all sorts of fun data things.
If you are interested in exploring, email me: MultiTechVisions@gmail.com
What can you do with it?
The idea is that you can create “chains” in Aigents that you can then call from your app through an API; these chains then collect data from your app and process it however you’ve configured the chain.
Here are the step types:
- Agent (call a LLM with a reusable instruction set)
- Branch Point (IF/THEN/ELSE type branching)
- Checkpoint (stops the run for
human in the loop) - Custom Instruction (call a LLM with a custom instruction set)
- External Data (retrieve a record from your app)
- Extract Table Data (send a bunch of text to LLM, it extracts data into specified table)
- JSON Extraction (send a bunch of text to LLM, it extract data into key/value pairings)
- Output (two types):
- Email (send an email using AppSheet’s email system)
- AppSheet (send data back to your app; new records and updates)
- Subroutine (include the steps from another chain in this one; alternatively, run a separate chain in parallel)
- Table Data (get multiple records from your app)
- Update Folder Variables (long-term memory system)
- Variable (make a variable for use down-stream)
Why not just use automation?
The real benefit is that I’ve solved a lot of the really fiddly technical hurdles of making something like this work. Most data in AppSheet apps is hierarchical, so it’s split up over many tables - so it’s not like we can get the data from one table and send it to the Ai… we need some way to put it all together and format it in a way that when we give it to the Ai it doesn’t get confused.
- And it turns out, there’s more fiddly bits involved with getting everything cleanly over to the Ai than one would hope for.
- Thankfully, I’ve solved those hurdles for you with this tool!
- Thankfully, I’ve solved those hurdles for you with this tool!
Another real benefit is that, if you’re doing things through Automation.. then that means that every time you want to update the chain in some way (add a new step, for instance), you have to change the app. Which means you might run into problems with versions (if you have users in the field), or if you break things… 
… you see what I’m saying.
- Using Aigents allows you to experiment freely
- If the chain is broken, who cares… it’s just the chain - your app is fine
It also allows for regular users, people that are non-technical to get there hands dirty and start building out Ai chains. I’m talking secretaries, mechanics, front desk personnel, people that you wouldn’t allow to get into the AppSheet editor because they’ll likely break something.
- These people can now freely experiment with building out Ai integrations
- And who better to make these sorts of IF/THEN/ELSE chains, than the actual people that do the work!