A platform for AI agents that handle long-running tasks: architecture and roadmap
The initial design of a platform for AI agents: persistent tasks, local and cloud execution, permissions and the first planned applications.
12 min read
AI solutions architect
From pilot to production, I design and build AI systems for document-heavy work across a variety of industries, from engineering and construction to finance.
Working across Australia and Japan
I'm Rounak Zoeb, a mechanical engineer turned data scientist turned AI solutions architect. I build AI applications for teams whose work runs on documents and data. I work with clients across cultures and languages, from the first requirements through to ongoing support.
Selected work
Bilingual AI platform for document translation and creation
A document translation and editing platform used by a Japan-based client, with editable outputs and review tools.
7 min read
doc-processing-toolkit
A Python framework for AI agents, with tools, persistent sessions and cost tracking across multiple model providers.
5 min read
Agentic research agent for capital projects data
A research tool that fills gaps in project records and gives analysts the sources they need to review each result.
6 min read
What I do
- AI applications
- Applications that use language models and tools to help people complete tasks, with evaluation, review and cost controls built in.
- Document processing
- Systems I design and develop for translating, editing and extracting information from PDFs, spreadsheets and other technical documents, and for drafting new documents from a range of source data.
- Data systems
- Data pipelines, data models, dashboards and predictive models, including the definitions and rules that AI agents work from, drawing on my work with engineering, project and financial data.
- Delivery and support
- Requirements, development, deployment and ongoing improvements, working directly with the people using the software.
How I work
I work directly with the people using the software, from understanding the task through to testing and support.
Understand the task
I start with the existing workflow: what people need to do, where it takes time and what a useful result would look like.
Build and test
I develop the application in stages, check it against realistic examples and review it with the people who will use it.
Improve it in use
After deployment, I work through feedback, investigate problems and keep the system up to date.
I use AI coding tools throughout development to prototype and iterate quickly, while staying responsible for the design, architecture, review and release of the final system.
How AI coding agents fit into my workflowInside a client project: a bilingual AI document platform
A scanned PDF arrives, in Japanese or English.
Each page is analysed and routed to a Word, Excel or image agent.
The team reviews each result beside the original page, and edits it directly or with the in-app editing agent.
Editable files are exported, with every page's cost logged.In production: Bilingual AI platform for document translation and creation
- A scanned PDF arrives, in Japanese or English.
- Each page is analysed and routed to a Word, Excel or image agent.
- The team reviews each result beside the original page, and edits it directly or with the in-app editing agent.
- Editable files are exported, with every page's cost logged.In production: Bilingual AI platform for document translation and creation