Here’s a video walkthrough: https://youtu.be/S4L3B21vXGY.
We initially set out to build a dev tool for API integration, and while building in the integration space, we realized existing workflow automation tools have a few key limitations. They still force users to do a lot of work like: Navigate through a sea of menus; Break down their workflow into discrete steps; Manually configure data transformations.
This led us to explore how we could make workflow automation way simpler by letting users describe what they want in plain English and having AI take care of the automation setup, replacing no-code GUIs or scripts.
Under the hood, DryMerge has two key components:
- A semantic layer that uses LLMs to interpret the user's request and map it to a series of pre-defined triggers and actions (we've built hundreds of these integrations).
- A data plane that orchestrates the actual execution, complete with smart field mapping, conditional logic, and human-in-the-loop checks.
When a user describes a workflow, our semantic layer generates multiple candidate plans, scores them, and selects the best based on prior successful/failed workflows. It extracts key entities and fields needed, and auto-generates a simple form for the user to fill in any missing details. Users can then iteratively describe, tweak, and test their workflow in the same chat.
The data plane then subscribes to the relevant event streams, executes the workflow steps, and handles gnarly aspects like pagination, retries, and rollbacks invisibly. We allow the semantic layer to delegate some values for runtime dependency injection from the data plane, to handle open-ended logic like classifying an email as urgent or summarizing a Google Meet transcript.
We integrate with 14 common services — we’d love for you to try it out and share what you think. Check it out at https://drymerge.com/app.
FYI: In good faith, I asked some simple javascript questions and stuff like "who is michael jordan" and got answers from the LLM. Perhaps adding some additional guardrails to ensure queries are workflow based could save you some tokens.
Great work!
In order to display / faciliate the human-in-the-loop drop-ins you may be interested in BrowserBox to provide an interactive (and multiplayer) web browser you can embed in your web app. You can check out a demo of it live here: https://browse.cloudtabs.net/signupless_session
1. For dynamic injection of arguments in your data plane, do you use LLMs?
2. What did you find you cannot do yet because of LLM limitations (and not because of lack of third-party integrations)?
3. I haven’t looked closely into your product, but is every “effect” of a workflow something that only the requesting user can see? Is this how you ensure bad or wrong things are not hurting other people or systems that are outside of user’s control?
2. Anything too complex (e.g. 5+ steps) tends to be unreliable. Also, any workflow where potential failure/unexpected behavior is too risky to leave up to an LLM.
3. The only actions we take are with our user's tools, so many workflows are simply organizing their information between their apps. However, e.g. gmails could be sent externally so we have guardrails/sanity checks to mitigate risk there.
What happens right now when the workflow fails mid-way? Do you ensure atomicity or durable execution?
I would appreciate 'whenever I post on X or Mastodon add this to a sheet and put text and URL in its own column'
Edit: Homepage link https://drymerge.com/