Zazmic’s New AI Agents: From SOW to Prototype to Code in 3 Days

The Fun, the Hard, and the Boring:
How We Killed the Worst Part of Product Development

Let’s face it: what comes in between an incredible client idea and actually building something cool is a good deal of boring admin work. All the project setup that keeps your team busy before they can even start coding. All the tedious paperwork. We know the pain.

At Zazmic, we meet hundreds of clients a year, which means we process over 300 Statements of Work (SOWs) annually—almost one a day! Our project managers used to spend two to three hours on a single SOW, painstakingly pulling information from meeting transcripts, spreadsheets, and presentations.

That amount of manual effort became a major drag on our productivity.

Below is how we solved it, slashing the path from a client meeting to deployable skeleton code to just three days.

Creating a Multi-Agent AI System—The Tech Under the Hood

As Zazmic’s CTO, Yann Kronberg, puts it in a recent podcast about these new agents, “We aren’t inherently a product company, so every tool we’ve built in-house has come from a necessity—automating tasks we do repeatedly every day.”

But we didn’t just want a quick fix. We wanted a solution that was smart, reliable, and integrated into our existing workflows. This transformation was achieved through a multi-agent system, where specialized models handle sequential, complex tasks that were once solely reliant on human effort. Here’s the technical architecture behind this pipeline:

Step One: SOW Automation

Once we’re wrapped up a client meeting, the journey begins immediately. The core technical challenge here is not just generating text from audio, but transforming unstructured transcript into a standardized, actionable Statement of Work (SOW), while controlling hallucination issues and discrepancies.

Zazmic’s solution employs a robust, modular architecture built around Google Drive integration:

• Intelligent Data Pre-processing: The system doesn’t dump all data into the LLM at once. Instead, a Python-based service first pre-processes all input files (e.g., transcripts, design files, estimate spreadsheets). It intelligently isolates only the necessary information, such as the “backlog” tab from the financial estimate or the executive summary from high-level documents.
• Modular LLM Architecture: To minimize latency and hallucination, the generation process uses a multi-step chain of separate LLM calls for each major section of the SOW (e.g., scope, deliverables, timeline, assumptions). This granular control ensures precise content for every part of the document.
• Context Aggregation & Grounding: A critical intermediary step uses an LLM to summarize each isolated data source into a unified knowledge base. This prepared context is used to ground all subsequent content generation, ensuring every statement is fact-checked against the client’s estimate. This process prevents scope creep by excluding concepts discussed in client calls but not formalized in the budget.
• Template-Based Generation: The final SOW is generated from a pre-defined Google Docs template. The service dynamically replaces designated placeholders (like ##client_name## or ##scope_summary##) with the LLM-generated, grounded content. This guarantees consistent, professional formatting, including proper handling of multi-level bullet points and bolding.

This process reduces SOW creation time from hours to just 30 minutes, allowing our product managers to focus on review and strategic alignment.

Step Two: Prototype Generation

With the requirements locked down in the SOW, the pipeline immediately pivots to visual design. This step requires the AI to interpret the structured project scope and translate it into a high-fidelity, clickable Figma prototype.

Within hours, the approved SOW is a visual reality in Figma, ready to serve as the definitive blueprint for development.

Step Three: Jira Ticket Orchestration

The third agent takes the approved visual design and the structured SOW to automate the entire project management backlog. The agent uses the following structure:

• Design and Scope Input: The agent ingests the structured SOW and the design data from Step 2.
• Advanced Planning Logic: The agent prompts Gemini to systematically decompose the project. It maps screens, components, and user flows into Epics, Features, and User Stories.
• Automated Backlog Creation: Vertex AI facilitates the secure and controlled calls to the Jira REST API. The agent handles the necessary API authentication and payload formatting to automatically create all tickets.
• Metadata Assignment and Criteria Generation: The agent automatically assigns crucial metadata (Component, Priority, Estimated Story Points) and generates detailed Acceptance Criteria. It embeds Google Drive deep links to the relevant SOW section and design files within each Jira ticket, ensuring 100% traceability.

The development team begins work on a perfectly structured Jira board, with all planning logic and execution managed by agents, saving days of manual project setup.

Step Four: Skeleton Code Generation

The final step delivers the immediate starting point for developers: a repository initialized with runnable, structurally sound skeleton code. The agents uses LLMs for parallel front-end and back-end scaffolding:

As a result, the development team is handed a secure, deployable skeleton code within just three days.

“What you see is a process that would take us six to eight weeks. Now, in three days, we have an SOW, a pretty good clickable prototype, features and epics in JIRA, and some skeleton code ready to be reviewed by our development team,” sums up Yann. 

What Happens to the Human Developer?

This AI Agent pipeline doesn’t just benefit project managers—it fundamentally changes the role of the developer.

For traditional engineering teams, productivity can go up by 10 to 100 times. The developer’s job shifts to becoming a reviewer and architect. They begin with skeleton code—often a pretty advanced set—and focus on building up the complex business logic, security, and integration, rather than boilerplate code.

This requires more analytical and reviewing skills, so more experienced developers will likely get the most immediate productivity boost. This change makes the job more engaging and interesting for the developer. Ultimately, everybody wins—the team, the customer, and the organization. 

Scaling Zazmic’s Product Development AI to Your Enterprise

Zazmic is taking this internal tool to the public, preparing to publish these agents on our marketplace and the Google marketplace by the end of 2025.

What teams can our pipeline fit? These agents are most effective for organizations that share a design-focused workflow, where all development is derived from a final design. For those who work differently, adoption requires training. While the SOW, UX/UI, and manufacturing aspects of the agent are robust, the engineering part is continuously being refined to fit diverse organizational workflows.

Customers now expect faster delivery, and our agents deliver.

Ready for AI? Let’s Build Your First Agent Together

The world of AI agents is just getting started. We’re on the verge of a productivity revolution where powerful AI tools handle the heavy lifting, saving your teams hours of manual work.

Many entrepreneurs understandably fear investing in AI and not seeing the immediate ROI. That’s where Zazmic comes in: we specialize in agile deployment, helping you quickly pivot to the solution that does work for your business.

As Zazmic’s CTO Yann Kronberg, who over his 20 years in tech & AI has made multiple successful startup investments and spinoffs, explained the mindset required in our recent podcast episode:

“There are a lot of things where you just don’t know if it’s going to work out right away. Ultimately, the success of any AI project hinges not just on tech, but on creativity and experimentation. You can do a lot more, and solve a lot more, with AI—that’s the fact. You just have to be brave and try a lot of different things.”

Ready to stop wasting weeks on work that could have been automated long ago and start building agents that generate real business value?