Build a tested Python package and FastAPI service that validates requests, calls a model, streams output, logs failures, and exposes documentation.

Compare hosted and open models for quality, latency, structured output, tool use, and cost, then implement configurable routing logic.

Create a versioned prompt service with Pydantic schemas, golden test cases, adversarial cases, automated evaluation, and release thresholds.

Ingest a mixed document corpus, apply metadata and access fields, index embeddings, and expose filtered semantic and hybrid search.

Build a cited RAG assistant with query rewriting, reranking, abstention, cache, tracing, permissions, and offline retrieval evaluation.

Prepare and validate a task dataset, run a small fine-tuning experiment, compare it with the baseline, and write a model decision memo.

Train a parameter-efficient adapter, track experiments, benchmark before and after behaviour, publish a model card, and document serving options.

Create a single agent with typed tools, explicit graph state, memory, retrieval, retries, budgets, stopping conditions, and test scenarios.

Build a persistent workflow with risk-based approval gates, escalation, audit events, recovery logic, and measurable intervention metrics.

Implement coordinator and specialist agents, then compare the result with a simpler single-agent baseline on quality, latency, reliability, and cost.

Build an MCP server and client workflow with scoped tools, authentication, secrets protection, prompt-injection tests, and a documented threat model.

Ship a deployed agent product with Python API, interface, RAG, tools, memory, approvals, evaluation, tracing, CI/CD, cost controls, runbook, and architecture case study.