Grounding
Curated a public career fact set and wrote instructions that keep answers tied to that material. When the facts do not support an answer, the assistant is directed to say so.
Project note AI assistant prototype
Building a career assistant that can explain the work.
I built and deployed a career assistant prototype as a small, practical AI product: a custom interface, a Cloudflare Worker, a curated set of public career facts, and DeepSeek R1 Distill running on Cloudflare Workers AI.
Explore the case studies01 / The work
What I built
Application
Built the visitor chat and a custom Cloudflare Pages Worker that validates requests, serves reviewed answers for common questions, and calls DeepSeek for other career questions.
Response handling
Added length checks, same-origin checks, low-temperature generation, and filtering for the model’s reasoning tags before an answer reaches the page.
Operations
Deployed the site and Worker on Cloudflare. An API rate limit, a model token cap, and the account’s free Workers AI quota help bound usage and cost.
02 / Technical clarity
Grounded, not trained.
DeepSeek R1 Distill is a pre-trained model. I did not train or fine-tune its weights. The technical work here is in choosing and curating the source facts, building the application around the model, handling its output, and operating it for real visitors.
The assistant can still make mistakes. Its answers are a guide to my public background, and the underlying résumé remains the source for role-specific details.
03 / Architecture
From question to answer
- VisitorAsks a career question
The site sends it to the chat endpoint.
- Cloudflare WorkerValidates and routes
Reviewed prompts get reviewed answers; other questions go to the model with approved facts.
- Workers AIRuns DeepSeek R1 Distill
The model generates a response from the provided context.
- WorkerChecks the output
It removes reasoning tags and returns the answer to the chat.