Hunch
An AI workflow orchestration platform that democratizes complex AI model interactions through a visual programming interface. By providing a spatial canvas for chaining AI models, it lets non-technical users create complex AI workflows without requiring coding expertise.
Hunch aims to simplify AI tool-building via a visual, no-code interface for chaining models. I briefly joined the project early on (2023) to contribute to the initial interaction design and explore the deeper potential for AI as a thinking partner, drawing on my background in systems design, tools for thought, and spatial interfaces.
Problem
Problem
As AI models become more powerful, combining them into useful workflows often remains a technical challenge requiring coding expertise. How can non-programmers chain together different AI capabilities (like text generation, image analysis, data extraction) to create custom tools or automations? Hunch‘s bet was that a visual, spatial canvas could make orchestration legible enough to be usable without turning everyone into a developer.
My work (early 2023)
My work (early 2023)
I joined for a short stretch in Hunch‘s early stage (2023), stepping in while my friend Ricardo Saavedra (who led design through early 2025) was on leave. The work was less “invent a new paradigm” and more “make the canvas readable when it gets real.”
Concretely, that meant:
| Area | What I changed | Why it mattered |
|---|---|---|
| Block types | Clarified block types so flows stay scannable as they grow | Reduced cognitive friction on large canvases |
| Taxonomy | Tightened the taxonomy (what a block is and how it behaves) | Made intent/behavior legible at a glance |
| Execution feedback | Improved execution feedback (state, progress, errors, outputs) | Helped users keep a stable mental model while the system runs |
Make the canvas scannable
One of the first challenges was helping users distinguish block types at a glance. I developed a modular color system for operations and restructured the block taxonomy so roles and capabilities were clearer (especially once flows became large and nested).
Make execution legible
The other pressure point was state: what’s running, what’s waiting, what failed, what produced output. I redesigned the context bar and refined the execution feedback loop to expose essential state changes without turning the interface into a dashboard.
Tensions
Tensions
We also explored a product-direction question that kept resurfacing: should Hunch be optimized for one-off automations, reusable “apps,” or something more emergent?
What Hunch chose to prioritize — modular tools, goal-oriented flows, user-defined macros — makes a lot of sense from a productization angle. But it leaves out a deeper layer of intelligence: systems that evolve. Feedback loops are how complex behavior emerges, how learning happens, how unexpected insights surface. Without them, you’re locked into linearity.
This divergence became formative for me. I was advocating for more open-ended collaboration between humans and AI, while the roadmap understandably focused on clearer, more market-ready outcomes like composable tooling. Both are valid paths — but they lead to different kinds of systems. Wrestling with that split directly informed the AI UX Paradigms talk I gave shortly after.
The path I hoped for
The path I hoped for
If Hunch had embraced feedback mechanisms from the beginning, it could have supported something closer to co-evolution between user and workflow: every execution becoming part of a growing pattern, not just an isolated run.
It would have opened the door to:
| Capability | What it supports |
|---|---|
| Persistent internal states | Workflows that evolve across executions instead of resetting |
| Looping logic | Pattern refinement rather than one-shot runs |
| Exploratory modes | Outcomes that steer future behavior |
| Fractal workflows | Tools that generate tools, not just results |
This wasn’t part of the product scope, but it anchored my thinking throughout.
Outcome (and how I use it)
Outcome (and how I use it)
My contributions (and Ricardo’s ongoing work) helped establish foundational UX patterns for the visual canvas: clarity, legibility, feedback. Hunch succeeds at making model chaining accessible through a no-code spatial interface.
Interestingly, I now use it primarily for the more open-ended exploration I was initially reaching for. The vast canvas, combined with free access to a wide range of models, makes it an excellent environment for developing thoughts visually, branching inquiries, and tracing the history of an idea spatially — uses perhaps not fully intended, but highlighting the power of the spatial paradigm itself.
While the deeper systemic ideas (feedback, adaptation) weren’t prioritized in the initial product, Hunch‘s success in democratizing model access and its flexible canvas demonstrate steps towards more fluid human-AI interaction.
The real influence — if any — was in challenging the team to consider what it means to build not just pipelines, but thinking environments.