Problem and intended user
Rubric: Does it work · Technical explanation
Who has this problem, what task are they trying to complete, and what does success look like for them in one or two sentences?
Technical Design Document Template
Every student completes this document for their individual capstone. It is the same instrument as the capstone assessment rubric: each section below maps directly to one or more rubric criteria, so what you write here is what reviewers use to verify your work.
Copy this page into your repo or export to PDF when complete. Replace each blank area with your project-specific content. The diagram in the architecture section should match your repository structure.
Scope and explicit non-goals
Rubric: Does it work · Architecture
List what the capstone will and will not do. Non-goals prevent scope creep and show reviewers you understand boundaries (e.g. no autonomous medical diagnosis, no unsupervised hiring decisions).
Architecture with diagram
Rubric: Architecture
Draw the system: user interface, API layer, retrieval or agent orchestration, vector store, LLM provider, MCP servers and external tools. Label data flow for a single request.
Data sources and handling
Rubric: Architecture · Documentation
Where does data come from (documents, APIs, databases)? How is it ingested, chunked, embedded, stored and refreshed? Note access controls on source material.
Retrieval or agent design and why that shape
Rubric: Architecture · Does it work
Describe your RAG pipeline (chunk size, reranking, hybrid search) or agent graph (tools, routing, memory). Explain why this shape fits the problem better than a single prompt or a different pattern.
Model selection with cost and latency reasoning
Rubric: Architecture · Technical explanation
Name the embedding model, LLM(s) and any fallback. Estimate cost per 1,000 queries and expected p95 latency. State what you would change if budget or latency requirements shifted.
Alternatives considered and rejected
Rubric: Technical explanation · Architecture
List at least two approaches you evaluated (e.g. fine-tuning vs RAG, single agent vs multi-agent, different vector stores or MCP vs direct API). Give a concrete reason each was rejected.
Evaluation approach and results
Rubric: Evaluation
Describe your golden dataset, automated harness (RAGAS, custom scripts or agent task benchmarks) and metrics. Paste or link summary results and note the weakest category.
Failure modes and behaviour under failure
Rubric: Reliability
What happens when retrieval returns nothing, the LLM rate-limits, a tool call fails or input is malformed? Document timeouts, retries, fallbacks and user-visible error messages.
Security, privacy and DPDP considerations
Rubric: Reliability · Documentation
How is PII handled? What data leaves your infrastructure? Note authentication, prompt-injection mitigations, audit logging and alignment with India's Digital Personal Data Protection Act where applicable.
Deployment
Rubric: Deployment
State the live URL, hosting platform, container image if used, environment variables required, CI/CD pipeline and health-check endpoint. Confirm the repo contains everything needed to reproduce the deployment.
Known limitations and next steps
Rubric: Technical explanation · Documentation
What does the system get wrong today? What would you build next if you had another month, tied to evaluation gaps, not feature wishlists.