Proma.ai Engineering

Lead engineering and product direction for Proma's AI assistant: a front-office worker that answers a business's customers, works for the owner, and builds the systems the business runs on.

I lead engineering for Proma.ai at Growthzilla, working with a remote US-based team. Proma is a platform businesses use to compose the systems they run on. The assistant is the product I lead on top of it, and it is where most of my work goes.

This case study covers only public information. It describes the engineering approach, not private metrics or internal implementation details.

What I’m building

The assistant is an AI front-office worker for a small business. Each business gets one, with one identity, one set of rules, and one memory of what the owner has told it. It changes how it behaves depending on who it is talking to and what is going on, but it is always the same assistant.

  • For customers, it answers on the business’s website from what it has learned about the business. When it cannot answer, it says so, takes a message, or brings in a person.
  • For the owner, it does the work and brings back only what the owner has to decide. The owner teaches it how the business runs in plain words, and it follows those instructions from then on.
  • For the business, it builds and runs the systems on Proma. The owner describes a form, a list, or an automation, and the assistant builds it and answers questions about the data.

One rule guides the product: the owner should feel they are putting the assistant to work, never doing the work themselves.

Architecture

Proma assistant architectureCustomer -> Website chat. Business owner -> Owner workspace. Website chat -> Assistant turn. Owner workspace -> Assistant turn. Assistant turn -> LLMs. Assistant turn -> State engine: allowed?. Assistant turn -> Proma systems: builds. Assistant turn -> Facts and rules: reads. Research -> Facts and rules. Facts and rules -> Owner workspace: needs youallowed?buildsreadsneeds youCloudflareKnowledge serviceCustomerLLMscalled each turnBusiness ownerWebsite chaton the business siteAssistant turnone per businessOwner workspacerooms, morning emailState enginewhat it may doProma systemsdatasets and formsResearchpublic pages to factsFacts and rulesand open questions

What I lead

  • Product direction for the assistant, working with the CEO and CTO
  • Its vision and architecture. I led the research into what the assistant should be and how to build it, and wrote the design the team builds against
  • Engineering across the full stack, from database design to deployment on Cloudflare’s edge platform
  • The rules for what the assistant may do. Code decides what is allowed; the model decides what to say
  • Knowledge and accuracy: learning a business from its public pages, and making sure no customer question it could not answer gets lost
  • Evaluation and releases: checking answers against what the business really offers before changes ship

Engineering focus

We learned early that we could not tool our way out of the hard problems. Each time the assistant got something wrong, adding another tool helped for a week. The real gaps were about state: what it already knew, what was still open, and what it had promised.

So the assistant is built like a worker, not a chat window over a list of tools. What it knows, what is waiting on the owner, and what it may do live outside the model, where code can check them. The model reads the situation and writes the reply.

That matters because the assistant speaks to customers in the business’s name. A confident wrong answer costs more than no answer.

The work crosses several layers:

  • Product layer: the customer chat, the owner’s workspace, and iteration speed
  • AI layer: LLM orchestration, tool use, prompt infrastructure, and structured outputs
  • Data layer: PostgreSQL-backed persistence, the business’s knowledge, and schema design
  • Infrastructure layer: Cloudflare deployment, edge runtime constraints, and production operations

Remote engineering

The team works remotely, so I optimize for clear written decisions, small reviewable changes, and enough documentation that future work does not depend on memory or meetings.

That matters more in AI product work because the product changes quickly. Clear writing keeps the system understandable while the team moves fast.

Reflection

Proma.ai is the kind of engineering I enjoy: shaping an AI product from idea to architecture to production, while keeping the system understandable enough for the team to keep evolving it.