Standard Template Constructs

Inspiration

Most code isn't new. We write the same pieces over and over, just with different names. AI tools can help, but most of them run online, cost money, and start from zero every time. We saw a huge problem in token consumption and sought to maximize efficiency in artificial intelligence and token usage. We wanted a tool that remembers what it's made and runs on your own computer.

What it does

You type what code you want in plain English. STC finds the closest match in its library of code templates, and a local AI fills in the blanks. If nothing matches, the AI makes a new template and saves it. The more you use it, the smarter it gets.

How we built it

We started by planning the whole project with Claude. Then we built the app in Python with Flask, and used a simple keyword search to find the best template. NVIDIA Nemotron runs on the user's own computer to fill in templates and make new ones. The app opens in its own desktop window with a clean, terminal-style look.

Challenges we ran into

Getting the AI to always answer in the right format was harder than we expected. It sometimes added extra text, so we wrote code to clean it up. We also had to find the right balance for matching, so STC didn't make new templates too often or too rarely. Making a local AI fast enough to feel smooth took some work too.

Accomplishments that we're proud of

STC works fully offline, with no internet, no API keys and no cost. It only uses the AI when it really needs to, so it stays fast. Its template library grows on its own over time. There's even a mode that works with no AI at all.

What we learned

Not every problem needs AI, because a simple search can do a lot of the work. Getting clean, reliable answers from an AI takes careful instructions. We also learned that the right-sized model often beats the biggest one.

What's next for Standard Template Constructs

Our next big step is showing people and companies how much STC can cut down on AI token usage. Most AI tools write every line of code from scratch, which burns through tokens and money. STC reuses templates, so the AI only fills in a few small blanks, using far fewer tokens for the same result. We'd also like to add a Nebius connection to give people with older laptops more power, and someday we'd love to turn STC into a VS Code extension.

Built With

• css
• flask
• flask-socketio
• gguf
• html
• javascript
• json
• llama.cpp
• lzma
• nvidia-nemotron
• openai-api
• pyqt6
• python
• qtwebengine
• scikit-learn
• tf-idf