AI Engineering Project Ideas

Build a terminal-based chatbot that maintains multi-turn conversation history using the OpenAI Chat Completions API.

• Accept user input in a loop and send messages to the API
• Maintain a rolling conversation history array
• Support a configurable system prompt via a config file
• Display streaming token output in real time
• Handle API errors and rate limits gracefully

Ingest a folder of text files, embed them with a sentence-transformer model, and serve a local search UI that returns the most relevant passages.

• Chunk documents into overlapping passages
• Generate embeddings using sentence-transformers
• Store vectors in a local FAISS or ChromaDB index
• Expose a simple FastAPI endpoint for query search
• Return top-5 results with similarity scores and source filenames

Create a Retrieval-Augmented Generation pipeline that lets users ask questions over a custom PDF knowledge base with cited answers.

• Parse and chunk PDFs using LangChain or LlamaIndex
• Embed chunks and store in a persistent vector store
• Retrieve top-k relevant chunks at query time
• Pass retrieved context to an LLM to generate a grounded answer
• Display source citations alongside each answer in a Streamlit UI

Fine-tune a pre-trained BERT-family model on a domain-specific sentiment dataset and deploy it as a REST API.

• Load and preprocess a labeled sentiment dataset using Hugging Face Datasets
• Fine-tune a distilBERT model with Hugging Face Trainer
• Track experiments and metrics with Weights & Biases
• Export the model and tokenizer to a local directory
• Serve predictions via a FastAPI endpoint with confidence scores

Build a LangChain agent that can search the web, read URLs, and write a structured research report on any given topic.

• Implement a ReAct-style agent with tool-calling support
• Integrate web search (SerpAPI or Tavily) and URL-scraping tools
• Allow the agent to plan multi-step research tasks
• Stream intermediate reasoning steps to the console
• Output a formatted Markdown report saved to disk

Build an automated evaluation harness that tests multiple LLM providers on a custom task dataset and produces a comparative scorecard.

• Define a JSONL dataset of prompts with reference answers
• Run each prompt against at least two models (e.g. GPT-4o and Claude)
• Score outputs using exact-match, ROUGE, and LLM-as-judge methods
• Aggregate results into a performance report with pass rates and latency
• Visualize results in a Jupyter notebook with comparative charts

Build a production-grade pipeline that accepts uploaded images, runs vision-language model analysis, and generates structured JSON reports stored in a database.

• Accept image uploads via a FastAPI endpoint
• Send images to a vision LLM (GPT-4o or LLaVA) with a structured output prompt
• Parse and validate the model response into a Pydantic schema
• Persist reports to a PostgreSQL database with image metadata
• Add async job queuing with Celery and Redis for scalability

Deploy a fully self-hosted RAG system using open-source models, a vector database, and a full observability stack for production monitoring.

• Serve an open-source LLM (Mistral or LLaMA) via Ollama or vLLM
• Use Qdrant as the vector store with collection management via API
• Instrument the pipeline with OpenTelemetry traces and LangSmith logging
• Implement hybrid search combining dense and sparse (BM25) retrieval
• Containerize all services with Docker Compose and document runbook

Build a multi-agent system where a planner agent decomposes coding tasks, executor agents write and run code, and a human approval step gates risky actions.

• Implement a planner agent that breaks a user task into subtasks
• Spawn specialized executor agents per subtask using LangGraph or AutoGen
• Execute generated code in an isolated Docker sandbox
• Gate file-write and shell-exec actions behind a human approval prompt
• Aggregate subtask outputs into a final deliverable with an audit log