New to coding
Start
Run Python, change a function, understand the data and make one test pass.
ECCU STEMNisa build · student workspace
Learn · build · prove
ECCU STEM technology field guide
Learn the tool. Know why it belongs.
A ranked map of the code, AI, data and delivery tools students can reuse across projects. Nisa is the anchor build; the engineering habits travel much further.
Choose your depth
Start simple. Open more detail when it helps.
Ready to contribute
Build
Own a small product outcome, work on a branch and prove it in the integrated app.
Ready for a challenge
Go deeper
Improve architecture, retrieval, evaluation, accessibility or failure handling.
Curious about professional systems
Look ahead
See how APIs, databases, deployment and governance grow from the same foundations.
Use now
The student stack is deliberately small
Python, uv, VS Code, Git, GitHub, Streamlit, Pydantic, JSON or YAML, pytest and Ruff are enough to build a real, safe first assistant. Codespaces is the recovery path when local setup gets in the way.
PythonuvVS CodeGit BashGitGitHubGitHub CodespacesStreamlitJSON and YAMLPydanticpytestRuffpython-dotenv
Full capability map
Open the six technology ranks
Optional detail for planning where the system can grow. A rank is earned through working evidence, not installed packages.
Full capability map
Open the six technology ranks
Optional detail for planning where the system can grow. A rank is earned through working evidence, not installed packages.
1Start here
Foundations
Write, run, explain and improve a small Python program.
- A program runs from one documented command
- Inputs are checked and errors are understandable
- The student can explain every function
Python · uv · VS Code · Git · GitHub
2Current build
Deterministic assistant
Build a useful assistant whose answers come from reviewed rules.
- A Streamlit chat interface works
- Knowledge is separated from application logic
- Answers show sources and safe boundaries
Streamlit · Pydantic · JSON/YAML · tests
3Next
Generative assistant
Use one language model without giving it control of the product.
- Provider calls pass through one gateway
- A fake model keeps tests deterministic
- Unsafe or uncertain requests fall back safely
Model gateway · HTTPX · structured outputs
4Advanced camp work
Grounded, tool-using assistant
Retrieve approved evidence and call small, typed tools safely.
- Claims link to retrieved evidence
- Tool arguments are validated
- Write actions require approval
Embeddings · retrieval · citations · typed tools
5Post-camp pathway
Production service
Turn the prototype into a service people can operate and trust.
- Conversations persist
- Authentication and permissions are enforced
- Failures, costs and model behaviour are observable
FastAPI · PostgreSQL · pgvector · deployment
6Only when justified
Institutional platform
Scale governance, workflows and channels around measured needs.
- The simpler architecture has a documented limit
- The new component has an owner and operating budget
- Recovery, audit and degraded operation are tested
Durable workflows · policy · multi-channel operations
Tool explorer
Find what you need
Showing 32 of 32 tools and tool groups.
Use nowRank 1 · Code
Python
The language used for Nisa’s behaviour, data work and tests.
- Why it belongs
- It keeps the first product in one language students can understand completely.
- Use it when
- Writing functions, validation, retrieval logic and integrations.
Use nowRank 1 · Code
uv
Creates the Python environment and installs locked dependencies.
- Why it belongs
- One tool handles project setup, packages and reproducible installs.
- Use it when
- Starting the project or adding a Python dependency.
Use nowRank 1 · Code
VS Code
The editor used to read, write, run and debug code.
- Why it belongs
- Its Python, terminal, Git and testing support live in one workspace.
- Use it when
- Working locally or inside GitHub Codespaces.
Use nowRank 1 · Code
Git Bash
Provides a consistent command-line environment on student laptops.
- Why it belongs
- Students can use the same project commands across Windows machines.
- Use it when
- Running uv, Git and project commands on Windows.
Use nowRank 1 · Quality
Git
Records changes to the project as a reviewable history.
- Why it belongs
- Small commits make experiments recoverable and teamwork understandable.
- Use it when
- Completing a coherent change that can be described clearly.
Use nowRank 1 · Quality
GitHub
Hosts the repository, issues, pull requests and team evidence.
- Why it belongs
- The project plan and the code change can be reviewed together.
- Use it when
- Claiming an issue, opening a pull request or reviewing a teammate.
Use nowRank 1 · Delivery
GitHub Codespaces
Provides a browser-based recovery workspace when a laptop setup fails.
- Why it belongs
- A setup problem should not cost a student an entire build session.
- Use it when
- Local installation is blocked or a clean environment is needed.
Use nowRank 2 · Interface
Streamlit
Turns Python functions into a visible chat application.
- Why it belongs
- Students get a useful interface without first learning React and browser state.
- Use it when
- Building the Nisa prototype and internal demonstrations.
- Hold back when
- A product needs a highly custom public interface or many independent frontend flows.
Use nowRank 2 · Data
JSON and YAML
Store reviewed FAQs, configuration and small structured datasets.
- Why it belongs
- They make the difference between program logic and product content visible.
- Use it when
- The data is small, reviewable and does not need concurrent updates.
- Hold back when
- Many users need to edit related records safely at the same time.
Use nowRank 2 · Quality
Pydantic
Defines and validates the shapes of requests, records and model output.
- Why it belongs
- Bad data is rejected at a clear boundary instead of failing later.
- Use it when
- Data crosses a file, API, model or tool boundary.
Use nowRank 2 · Quality
pytest
Checks expected behaviour automatically.
- Why it belongs
- Tests make refactoring safer and turn requirements into executable evidence.
- Use it when
- Adding a behaviour, repairing a bug or guarding a safety rule.
Use nowRank 2 · Quality
Ruff
Finds common Python mistakes and formats code consistently.
- Why it belongs
- The team spends less time debating style and more time reading behaviour.
- Use it when
- Before every pull request and in continuous integration.
Use nowRank 2 · Delivery
python-dotenv
Loads local environment settings without putting secrets in code.
- Why it belongs
- Configuration changes by environment; credentials must never enter Git.
- Use it when
- A local app needs a development-only setting or API key.
Learn nextRank 3 · Interface
HTTPX
Makes typed, timeout-controlled HTTP requests from Python.
- Why it belongs
- External model and data providers are network dependencies that can fail.
- Use it when
- Calling the selected model provider or a reviewed public API.
Learn nextRank 3 · AI
One model gateway
Keeps provider-specific code behind a small application interface.
- Why it belongs
- Nisa can change providers later without spreading vendor code everywhere.
- Use it when
- The deterministic assistant and its safety tests are already reliable.
- Hold back when
- Students cannot yet explain the non-AI version of the same feature.
Learn nextRank 3 · AI
Fake model
Returns predictable responses during tests without network calls or cost.
- Why it belongs
- Most tests should verify our code, not gamble on a changing model response.
- Use it when
- Testing prompts, fallbacks, usage records and failure handling.
Optional labRank 4 · AI
Pydantic AI
Defines typed agent dependencies, outputs and tool calls.
- Why it belongs
- It becomes useful after students understand direct model calls and ordinary functions.
- Use it when
- Nisa needs multiple validated tools or structured model results.
- Hold back when
- A direct provider call and two ordinary functions remain easier to explain.
Optional labRank 4 · AI
RapidFuzz or BM25
Retrieves relevant text with understandable keyword-based methods.
- Why it belongs
- A simple search baseline reveals whether embeddings actually improve results.
- Use it when
- Building the first retrieval evaluation set.
Optional labRank 4 · Data
pypdf / pdfplumber
Extracts text and page provenance from approved PDF documents.
- Why it belongs
- Reliable citations require preserving the document and page that support a claim.
- Use it when
- The approved knowledge base includes text-based PDFs.
- Hold back when
- Scanned pages require OCR or complex layouts require a stronger parser.
Optional labRank 4 · AI
sentence-transformers
Turns text into vectors for meaning-based retrieval experiments.
- Why it belongs
- Local embeddings can support a controlled comparison with keyword search.
- Use it when
- A labelled evaluation set shows keyword retrieval misses semantic matches.
Optional labRank 4 · Data
Chroma
Stores vectors for a small learning experiment.
- Why it belongs
- It offers a quick laboratory before adopting production database infrastructure.
- Use it when
- Students are comparing retrieval approaches on a small approved corpus.
- Hold back when
- The product already needs durable relational data, permissions and operations.
Learn nextRank 5 · Interface
FastAPI
Exposes the Python application through validated HTTP endpoints.
- Why it belongs
- It separates the user interface from the backend when the prototype grows.
- Use it when
- Another interface, integration or client needs to call Nisa.
- Hold back when
- The Streamlit prototype is still the only client and teaches the requirement clearly.
Learn nextRank 5 · Data
SQLite and SQL
Teach durable, queryable conversation and feedback storage.
- Why it belongs
- Students should understand SQL before an object-relational mapper hides it.
- Use it when
- Conversations must survive an application restart.
Add on evidenceRank 5 · Data
PostgreSQL + pgvector
Stores application records and embeddings in one production database.
- Why it belongs
- Transactions, joins, permissions and vector search can share one operational system.
- Use it when
- The deployed service needs multiple users, durable retrieval and backups.
- Hold back when
- A classroom prototype is still well served by files or SQLite.
Add on evidenceRank 5 · Data
SQLAlchemy + Alembic
Maps application records and versions database schema changes.
- Why it belongs
- Production data structures need deliberate, repeatable migrations.
- Use it when
- PostgreSQL becomes part of a maintained service.
Add on evidenceRank 5 · Interface
Next.js + TypeScript
Builds a fully custom web interface around the Python service.
- Why it belongs
- It provides control over routing, browser state and interaction design.
- Use it when
- A tested user need cannot be served well by Streamlit.
- Hold back when
- It would make beginners learn two languages before the product works.
Add on evidenceRank 5 · Delivery
Docker Compose
Runs a small multi-service development environment consistently.
- Why it belongs
- It becomes valuable when the API, database and supporting services must start together.
- Use it when
- The system has several services with reproducible local setup needs.
- Hold back when
- Students only need Python and Streamlit to run the current product.
Learn nextRank 5 · Delivery
GitHub Actions
Runs quality checks automatically for every shared change.
- Why it belongs
- The repository can enforce the same Ruff and pytest checks for everyone.
- Use it when
- The team is merging pull requests into a shared branch.
Add on evidenceRank 5 · Delivery
Structured logs, Sentry and Langfuse
Makes application failures and AI behaviour observable.
- Why it belongs
- Operators need to trace errors, retrieval, tool calls, latency and cost.
- Use it when
- Real users depend on the deployed service.
- Hold back when
- The team lacks the capacity to review and act on the collected signals.
Add on evidenceRank 6 · Delivery
Redis + Celery
Runs retriable background work outside the web request.
- Why it belongs
- Long document processing should survive API restarts and scale independently.
- Use it when
- Measured jobs are slow, concurrent, retriable or must survive restarts.
- Hold back when
- FastAPI background tasks or synchronous processing remain reliable enough.
Not yetRank 6 · AI
LangGraph or Temporal
Coordinates branching or durable long-running workflows.
- Why it belongs
- These solve real workflow problems, but also create a new operating subsystem.
- Use it when
- A measured workflow requires checkpoints, human pauses or durable recovery.
- Hold back when
- One agent with controlled tools can complete the task.
Not yetRank 6 · Delivery
Kubernetes, Kafka and multi-agent platforms
Operate large distributed systems with specialised scaling needs.
- Why it belongs
- They belong after product demand demonstrates the problems they solve.
- Use it when
- Multiple independently scaled services have dedicated operators and budgets.
- Hold back when
- A modular monolith and one database can still serve the product.
Safety is part of the stack
- Authority comes before retrieval, model output or tool use.
- Use only reviewed public NIS information in the student build.
- Never collect or expose a person’s account, entitlement or identifying data.
- Every supported factual answer should show its source.
- When evidence is missing, say so and direct the user to an official channel.
- The model proposes; ordinary code validates; the user approves consequential actions.
- Secrets stay out of source code, browser code, screenshots and chat messages.
- Most automated tests use fakes instead of paid or changing model responses.
Turn knowledge into evidence