Build Claude-Powered NLP Foundation

Claude will be the backbone of a new natural-language-processing system that I’m putting in place for my product. The initial scope centres on three capabilities: reliable text classification, accurate sentiment analysis, and a conversational chatbot that can surface answers from our own content.

My training corpus is ready: a large stream of social-media posts plus SEO-focused text collected from our marketing team. I can supply it in JSON or CSV along with basic labels where they already exist. You’ll need to clean, augment, and align this data so it fits neatly into Claude’s fine-tuning workflow (or an equivalent retrieval-augmented approach if that proves more efficient).

Here’s what I expect as concrete deliverables:
• An end-to-end Claude-compatible training pipeline (data prep, experiment tracking, version control).
• Three fine-tuned models or prompt templates covering classification, sentiment, and dialogue, each exposed through a simple REST or Python API.
• Short demo scripts that show the system responding to fresh inputs.
• A concise deployment guide so my internal developers can reproduce the environment on AWS or GCP.

Acceptance criteria: every model should reach at least 85 % macro-F1 on our hold-out set, and the chatbot must answer at least 8 of 10 predefined queries correctly during a live test.

If you have recent experience with Claude’s SDK, LangChain, or similar orchestration tools, this will move fast. Feel free to suggest additional metrics or workflow improvements—clarity, reproducibility, and clean code matter more to me than flashy dashboards.