Inspiration
DecenAI came from a simple frustration: job, scholarship, and application workflows are scattered across too many sites, require too much repetitive clicking, and break easily when a user tries to automate them with a naive script. The project was inspired by the idea of a visible, local, approval based desktop agent that can search, rank, draft, and assist without acting like a black box.
What it does
DecenAI is a local desktop automation agent that helps users search for internships and scholarships, keep structured results in SQLite, rank opportunities with a weighted scoring engine, open useful browser tabs, draft research artifacts, and support live control through Telegram, voice, and a local dashboard. It uses a planner plus tools architecture, where the model decides the next step and the browser, desktop, memory, and approval layers carry out the work.
How we built it
We built DecenAI as a hybrid local operator. A local model runs through llama.cpp as the planner, FastAPI serves as the orchestration layer, Playwright handles visible browser automation, pywinauto and Windows UI Automation handle desktop control, SQLite stores memory and artifacts, whisper.cpp powers local speech to text, Piper powers local text to speech, and Telegram acts as a remote control and approvals channel. We also added structured opportunity models, ranking, approvals, document tracking, and a recipe system for internship and scholarship search.
Challenges we ran into
The biggest challenge was making the agent act intentionally instead of literally. Early behavior showed that raw sentence queries could send the browser into bad searches, including Google Maps pages and junk links instead of real internship results. Another major challenge was balancing local model limitations with useful automation, which pushed us toward compact state, structured recipes, validation, and approval checkpoints instead of unconstrained autonomy.
Accomplishments that we’re proud of
We are proud that DecenAI became more than a script. It now has typed opportunity models, a real weighted ranking engine, persistent SQLite memory, voice and Telegram control, document tracking, and a planner plus tool architecture that can be extended into internships, scholarships, desktop actions, and browser workflows. We are also proud that the system is local first and visible, which makes it much easier to trust, debug, and demonstrate.
What we learned
We learned that good automation is less about making the model “smarter” and more about giving it structure. A constrained planner, strong validation, persistent memory, and recipe based execution work much better than dumping full user instructions into a search box. We also learned that local multimodal style assistants are most useful when the model plans and the tools execute.
What’s next for DecenAI
Next, we want to make DecenAI more reliable across real world application flows. That includes stronger site specific recipes, better selector memory, richer application queues, tighter resume and cover letter tailoring, more polished live dashboard controls, and broader support for end to end desktop workflows. The long term goal is to turn it into a general personal operator that can search, organize, draft, and act across both the web and the desktop while still staying local, visible, and approval aware.
Log in or sign up for Devpost to join the conversation.