I picked these because a basic RAG chatbot or simple agent is already on every CV, and certificates only teach theory. Each project here is built around a problem AI teams actually have in production: evals, cost, reliability, security and scale. Every one is small enough to finish in a weekend and gives you something real to talk about in an interview.
Put a small model behind an API, feed it live data, and get an alert the moment its inputs stop looking like its training data.
You can keep a model healthy after launch, which is what most MLOps interviews are really about.
Train a model in CI, log it, and promote it to production only when it beats the model already there.
You understand experiment tracking, a model registry and CI/CD for models as one flow.
Take one model out of a notebook and ship it behind an endpoint that scales up under load and back down when quiet.
You can turn data science work into a service other teams can depend on.
Build a scheduled feature pipeline, store the features, and serve a model that reads the same features at training and at prediction time.
You can stop the most common silent bug in ML systems: features that differ between training and serving.
Search a real set of documents with hybrid search and reranking, show where every answer came from, and score how often it finds the right document.
You can build retrieval that a company could trust, and prove it with a number instead of a demo.
Write a test suite for an LLM feature that runs on every change and blocks the release when answer quality drops.
You treat LLM features like real software, which is exactly what teams struggle with.
Put a gateway in front of several models, send simple requests to a cheaper one, and switch provider automatically when one goes down.
You can cut LLM costs and keep a product up during a provider outage.
Serve an open source model yourself and measure tokens per second, latency and what each request really costs next to a paid API.
You can answer the build or buy question for models with real numbers.
Build an agent that reads documents and still behaves when one of them hides an instruction telling it to do something else.
You think about security from the start, which few AI engineers can show.
Build an agent that searches, calls an API and writes to a database, and stops for a human to approve anything risky.
You can build agents a business would actually let run.
Run two workloads on one GPU with time-slicing, add quotas per team, and show the cost saving.
You can run GPU infrastructure, not just use it. Very few candidates can show this.
Add tests to a pipeline so bad data stops before it reaches a model or a dashboard, with an alert explaining what failed.
You protect everything downstream, which is the job of a data platform engineer.
Want the full spec for one of these, with the architecture and the steps? Send me a DM with the project name.
Send me the project name and I will send you the full spec, plus which one I would build first for your background.
Message me on InstagramOpens an Instagram DM