Refront: How We Built AI Workflow Automation

Introduction

When we started AVARC Solutions, we kept building similar automation layers for different clients: ingest data, classify it with AI, route it to the right person, and trigger follow-up actions. After the third time we rebuilt the same pipeline from scratch, we knew it was time to productize it.

That product became Refront — an AI-native workflow automation platform that lets teams design, deploy, and monitor intelligent automations without stitching together a dozen SaaS tools. In this post we share the technical journey from idea to production.

The Problem Refront Solves

Modern businesses run on workflows. A support ticket arrives, needs to be categorized, assigned to the right agent, and answered within an SLA. A new lead fills out a form and needs scoring, enrichment, and a personalized follow-up. An invoice is received and needs matching to a purchase order before approval.

Most companies solve these with a mix of Zapier, Make, custom scripts, and manual work. The result is fragile, hard to debug, and impossible to scale. Refront replaces that patchwork with a single platform where AI handles the decision-making and orchestration.

The key differentiator is that Refront treats AI as a first-class citizen in the workflow. Instead of bolting an LLM call onto a traditional automation, the entire routing and classification logic is driven by language models that improve over time.

Architecture and Technology Stack

Refront is built on a modern event-driven architecture. At its core is a workflow engine that processes events through a directed acyclic graph (DAG) of steps. Each step can be a transformation, an AI inference call, a conditional branch, or an external API call.

The backend runs on Node.js with Supabase for data persistence and real-time subscriptions. We chose Supabase because it gave us Postgres, authentication, storage, and real-time out of the box — all things we needed without wanting to manage them ourselves.

For the AI layer, we built an abstraction that supports multiple providers — OpenAI, Anthropic, and Google — so customers are never locked into a single model. The abstraction handles retries, fallbacks, and cost tracking transparently.

The frontend is a Next.js application with a visual workflow builder. Users drag and drop steps, configure AI prompts, and connect external services. Every change is versioned so teams can roll back without fear.

Building the AI Classification Engine

The heart of Refront is its classification engine. When an event enters a workflow — say, an incoming email — the engine uses a fine-tuned model to determine intent, extract entities, and decide the next step.

We started with prompt engineering using GPT-4 and quickly learned that generic prompts produce inconsistent results at scale. So we built a feedback loop: every time a human corrects a classification, that correction feeds into a training set for the next model iteration.

After three months of production data we saw classification accuracy jump from 78 percent to 94 percent. The remaining six percent are edge cases that get routed to a human reviewer, which in turn generates more training data. It is a virtuous cycle.

Lessons Learned in Production

Running an AI-powered platform in production teaches you things that no amount of prototyping can. The biggest lesson was about latency. Users expect automations to feel instant, but an LLM call can take two to five seconds. We solved this by running classification asynchronously and showing a "processing" state in the UI.

The second lesson was about cost. LLM inference is not free, and a customer running 50,000 events per month can accumulate significant API costs. We built a caching layer that deduplicates identical inputs and a model-routing system that sends simple tasks to cheaper models.

The third lesson was about observability. When an automation fails at step seven of a twelve-step workflow, the user needs to know exactly what happened. We invested heavily in structured logging and a visual debugger that shows the state of every variable at every step.

Conclusion

Building Refront taught us that the future of automation is not about connecting APIs — it is about giving AI the context it needs to make good decisions. The platform is now in production with multiple clients and handles tens of thousands of events per week.

If you are drowning in manual workflows or outgrowing your current automation tools, reach out to us. We would love to show you what Refront can do for your team.