HoneyFlows

hack4ai-goals
honey-flows-data-processing-idea
Demo of a flow to identify if client needs coaching or mental health help.
BeeWellAI team Logo
HonewFlows Tool Logo
HonewFlows Logo

Inspiration

Didem and I aim to streamline matching in our WellBeeAI project (https://wellbeeai.com/) . A former colleague suggested exploring AI's potential at a hack4ai event. Our initial plan: a versatile 'lego' approach, creating an adaptable 'tinder for everything' tool, saving months of coding."

What it does

The solution simplifies building complicated LLM-based data processing flows using low-code visual tools.

Use cases examples:

matching: linking together coaches/psychotherapists with customers and vice-versa
client-routing: identifying the customer groups based on text inputs, e.g. for Orbit8
complex automations: making posts in LinkedIn, Slack automation etc
easier prompts testing: stop copy/pasting prompts and prompts that depend on them, build a chain you need and call it as many times as needed

How we built it

Core stack and technologies:

TypeScript
NodeJS
Node-Red IoT platform
Express web-server

3rd party tools:

OpenAI REST API

Testing tools:

LM Studio for local llama debugging
Insomnium REST Client (https://github.com/ArchGPT/insomnium) for testing purposes

Challenges we ran into

How to guarantee the quality of the prompt-based chains.

Accomplishments that we're proud of

What we learned

Thanks to my ex-colleague Filipe whom I met in hackathon, I got an idea how to ensure prompt quality with 3rd party LLM APIs. Over time, user inputs may degrade outputs. By collecting prompt data in production, we can automatically validate it monthly. If quality declines, we can auto-tune prompts with AI itself until restored.

What's next for HoneyFlows

Convert PoC to an MVP:

Finalise the solution so anyone could use it after buying a subscription plan
Dogfooding and plugging the engine in our WellBeeAI product (https://wellbeeai.com/)
Partner with Orbit8 team (https://orbit8.app/), so they could use our engine in their solution