AI/ML Engineering Projects
Every project below is a real AI/ML problem teams hit in production - flaky LLM integrations, runaway cost, weak retrieval, no evals. Fix it in a live cloud workspace and it lands on your portfolio as a clickable proof point.
22 projects to build
Restore a Broken LLM API IntegrationRestored the LLM API integration by correcting the endpoint path and adding the required model field to the reHarden an LLM Pipeline Against API FailuresWrapped LLM API calls in retry + try/except so the batch pipeline survives transient failures and returns partParse JSON From an LLM (Strip Markdown Fences)Stripped the ```json fences before json.loads() so the extractor reliably parses LLM responses that ship insidBuild a Text Summarization Endpoint With an LLMWired the /summarize endpoint to the LLM with a tight prompt and response parsing so the API returns shorter tAdd Conversation Memory to a ChatbotForwarded the full conversation history to the LLM and capped it with a sliding window so the bot remembers coRestore Token Streaming on a Chat EndpointSet stream=True on the LLM call and switched the endpoint to a Flask generator with chunked transfer encoding Cache LLM Responses in RedisAdded a Redis cache keyed on the prompt with 5-min TTL so identical questions return instantly instead of rounBuild an LLM Support Ticket ClassifierShipped a CLI classifier that routes support messages into billing/technical/general using a constrained LLM pConnect a Chatbot to AWS BedrockFilled in chat() to call the OpenAI-compat LLM endpoint (mock-llm on the shared service - same wire format as Undo a Git Commit With git reset --softUsed `git reset --soft HEAD~1` to roll the last commit back into the staging area, preserving the file edits wGenerate Text Embeddings With OpenAIFilled in embed_docs() to call the embeddings endpoint (nomic-embed-text, same shape as OpenAI's embeddings APBuild an LLM Content Moderation ServiceStood up a FastAPI moderation service backed by an LLM classifier with safe/warning/blocked tiers, JSON outputCut OpenAI Costs by Caching With RedisWrapped the LLM call in a Redis cache-aside layer keyed on the prompt (24h TTL) so repeat questions skip the LPower RAG Search With a Vector DatabaseBuilt a RAG search loop - embedded each doc with nomic-embed-text, indexed into ChromaDB, then for each query Fix the Chunking That's Wrecking RAG RetrievalReplaced a character-slicing chunker that shredded docs into meaningless fragments with sentence-aware chunkinRe-Rank RAG Results So the Best Answer Lands on TopAdded a re-ranking stage that re-scores the retriever's top-k candidates with a relevance function and reorderMake an LLM Pick the Right Tool (Function Calling)Wired the function-calling loop - asked the LLM to choose a tool as JSON, parsed that choice, and dispatched tBuild an End-to-End RAG PipelineWired chunk → embed (TF-IDF) → retrieve (cosine top-k) → augment → generate (mock-llm) into a single `answer(qBuild an LLM Evaluation Harness (Golden Set)Wrote a generic evaluation harness that runs an injectable LLM function against a golden set, scores each respRoute LLM Requests to Cheaper ModelsRouted short prompts (<80 chars) to mock-llm (cheap) and longer ones to llama-guard3:1b (more capable), so simBuild a Tool-Calling Agent LoopImplemented the agent loop - call the LLM, run any ACTION it requests, feed the result back as an Observation,Stream a RAG Answer With Inline CitationsTurned a buffered RAG endpoint into a streaming one - retrieve the context, stream the LLM answer token-by-tok
Plus 16+ more AI/ML projects in the workspace, with new ones added regularly.
More for AI/ML engineers
Build your AI/ML portfolio free. Fix real systems in a live cloud workspace - every fix is yours to keep.
Start free →