Jan Zheng

Jan Zheng is a designer and developer whose AI work includes Fresh, an invoicing app he built for himself, and the Groq MCP Server, for which Groq credits him. His account of building Fresh emphasizes concrete decisions: introducing AI for document extraction, reviewing its outputs, and restructuring generated code by hand.[2][3]

Zheng built Fresh because existing invoice tools did not fit his consulting workflow: clients used different payment methods, while maintaining separate Google Docs templates made invoices tedious to create and track. In his March 13, 2025 article, “Vibe coding an invoice app”, he described a tool with editable invoices, granular undo and redo, PDF previews and payment links. Its first few versions contained no AI features.[2]

He added AI when converting Word and PDF documents into the app’s invoice schema. An LLM extracted invoice data into JSON; a vision model handled OCR for PDFs containing images rather than text. He used Groq-hosted Llama models and justified choosing inexpensive OCR because he would review every output: “The OCR is just there to fill in the template for me.” His report of encountering no extraction errors came with an explicit qualification: a tiny sample and no evaluations.[2]

Fresh also changed his judgment about chat interfaces. Zheng found chat useful for showing what the AI had changed and why, displaying generated key/value changes, and accepting corrections or updates. He could discuss line items and ask for estimates from a rough description, rather than entering every change through form fields.[2]

During development, Zheng rewrote Fresh from React to HTMX and then to Hono and Alpine.js. He preferred separating HTML templates, client-side UI logic and server functions because it let him give a coding model the portions relevant to a particular feature. For animation work, for example, he could select the templates and corresponding Alpine functions; for chat, he could focus on server handlers and the client functions calling them. He regarded this separation as easier to manage for a small personal app than supplying large amounts of mixed layout and logic.[2]

Zheng did not describe Fresh as an autonomously generated application. He manually refactored it, rejected Claude’s proposed changes and fixed bugs himself. AI assistance helped him attempt unfamiliar tools and features, including Canvas-based PDF previews, but he argued that periodic restructuring remained essential. In his words, “AI has a tendency to patch mistakes with layers and layers of duct tape, never fixing the leaks themselves.”[2]

Groq’s official examples page credits Zheng for the Groq MCP Server, which lets Claude and other Model Context Protocol clients query models hosted on Groq. The project repository documents tasks such as extracting image information as JSON, generating speech, transcribing and translating audio, and processing batches of prompts. It also documents agentic tasks using Groq’s compound-beta system and acknowledges the ElevenLabs MCP Server as inspiration.[3][4]

Before these projects, Zheng’s work combined interface design, coding and tools for a scientific community.[6]

In the Builders Club interview, written collaboratively by Zheng, Rebecca Williams and Annie Liao, Zheng recalled studying computer science and psychology before pursuing a master’s in human-computer interaction at Carnegie Mellon. He wanted to apply “game design techniques” to difficult problems. He subsequently worked at agencies on campaigns for companies including Coca-Cola and L’Oréal, then moved into startup work at Ubiquiti.[6]

Zheng recalled meeting Jessica Sacher at a swing dance and starting Phage Directory with her. He described it as a community platform that finds bacteriophages for patients. His Phage Directory author biography describes his contributions as designing and building tools and helping write and organize the newsletter Capsid & Tail.[6][5]

In December 2023, Zheng expressed enthusiasm for practical AI-assisted development, particularly asking Cursor to write code, rather than predictions that AGI was near. He argued that biology and chemistry offered more interesting opportunities than copying existing SaaS products: “Using AI to disrupt SaaS apps is like building a flashlight app on the first iPhone.” He also described spending much of his working week prototyping through design and code, favoring familiar tools while experimenting selectively with new ones.[6]

Zheng appeared in Latent Space’s first-anniversary listener interviews, published on March 9, 2024. The publisher listed his segment at 01:39:25 with the descriptor “Australia, AI UX.”[1]