Offline Coding Assistant on Consumer Hardware
Cloud coding assistants send your source code to somebody else's server, which rules them out for anyone working under a confidentiality agreement. Build a coding assistant that runs entirely on a normal laptop using a small open model, and measure honestly how much capability that costs. The contribution is the measurement: a reproducible benchmark of local models against cloud ones on the same tasks, on hardware students actually own.
Advanced
Needs a team, real planning, and a supervisor who knows the area.1 semester · 2 students
7 to ship
3 optional extrasSuggested stack
What you should ship
- An editor extension that provides code completion and explanation from a locally running model
- A retrieval layer that pulls relevant files from the open project into the prompt, using a code aware parser rather than plain text splitting
- Support for at least three open models of different sizes, switchable at runtime
- A benchmark of at least 50 tasks drawn from real repositories, scored automatically
- Measurements of latency, memory use and battery impact for each model on the same machine
- A comparison against one cloud model on identical tasks, with the gap quantified
- A written recommendation on which model size is worth running locally and for what work
If you have time left
- Quantisation experiments showing the quality cost at each precision level
- A multi step mode where the assistant can read several files before answering
- Verification that no network request leaves the machine during operation
The problem
A local model keeps your source code on your machine, costs nothing per token, works on a plane, and has no usage quota. It is also smaller and less capable than what a cloud provider will run for you.
Everybody states that trade off. Almost nobody measures it on the hardware a student or a small company actually owns.
What makes this a real project
The assistant is the easy half. The research contribution is the benchmark.
- Task selection. Fifty tasks that are actually representative is harder than it sounds. Pull them from real repositories with known correct commits, so scoring can be automatic rather than a judgement call.
- Fair comparison. Same prompts, same retrieved context, same scoring, one variable changed at a time. This is where most comparisons of this kind fall apart.
- Context retrieval. Splitting source code every 500 characters cuts functions in half. Use a parser that understands syntax so you retrieve whole definitions.
- Honest reporting. If the local model is much worse, say so and quantify it. A well measured negative result is a good dissertation.
Suggested approach
Build the benchmark harness first, before the extension. Run it against a cloud model to establish your ceiling. Then build the local pipeline and see how close you get.
Fix one machine as your reference hardware and report its specification. Numbers without hardware context mean nothing.
How you will know it worked
Somebody else can clone your repository, run one command, and reproduce your table on their own machine. That is the bar.
Ideas and guidance, not finished projects
These are project ideas and scoping guidance, published free for students to use as a starting point. I do not build, write, or sell final-year projects, and I do not complete coursework for anyone. Take an idea, make it yours, and build it.
Not sure this one fits you?
I help students pick an idea that matches their skills and their deadline, then scope it down until it is finishable. Guidance only, never done for you.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Every session is free; a few slots open each week.
Follow along
New writing, resources and project ideas land here first.