8000 GitHub - duboc/gemini-dev-demos · GitHub
[go: up one dir, main page]
Skip to content

About

No description, website, or topics provided.

Resources

Readme
Apache-2.0 license

Contributing

Contributing
Activity

Stars

5 stars

Watchers

6 watching

Forks

0 forks
Report repository
BranchesTags

Repository files navigation

Gemini developer lifecycle demos

This repository contains 13 interactive Streamlit demos that call Google's Gemini models through Vertex AI. Each demo applies a Gemini model to one stage of the software development lifecycle: inspecting a codebase, migrating legacy code, generating test scripts from screen recordings, reviewing user experience and accessibility, turning user stories into code, and generating data pipelines.

The demos target developers, solution architects, and technical presenters who want runnable examples of AI-assisted engineering. Treat the code as demonstration material that you read, run, and adapt, not as production software.

Animated Gemini logo

Before you begin

Install and configure the following:

  • Python 3.12. The container image in Dockerfile builds on python:3.12, so use the same version locally to match it.
  • A Google Cloud project with billing enabled.
  • The Google Cloud CLI, authenticated against that project.
  • Docker, if you plan to build the container image or deploy to Cloud Run.

Enable these APIs in your project. The setup.sh script enables all of them for you:

  • aiplatform.googleapis.com
  • cloudbuild.googleapis.com
  • run.googleapis.com
  • artifactregistry.googleapis.com
  • iam.googleapis.com
  • storage-api.googleapis.com

Gemini model identifiers

utils_vertex.py holds every Gemini identifier this repository sends to Vertex AI, in two module-level constants. No file under apps/ carries a model literal of its own: every demo imports MODEL_ID, and a demo that offers a model picker lists that one identifier, so a single edit reaches all 13 demos.

Identifier Constant in utils_vertex.py Availability on Vertex AI
gemini-3.8-flash MODEL_ID Listed under "Models available for shorter availability periods", a tier whose members retire 45 days after Google ships a replacement. The page states no fixed date.
gemini-embedding-001 EMBEDDING_MODEL_ID Listed among the embeddings models Vertex AI serves. No demo calls it, so the constant names the model an embedding feature would reach for.

Both identifiers come from Model versions and lifecycle, the page that states which models Vertex AI serves and for how long. Read it again before you trust either one. A 45-day window is short enough that a checkout a few months old can name a model the API no longer answers, and Vertex AI responds to a retired identifier with a not-found error rather than a fallback.

The same page dates the end of service for gemini-embedding-001 no sooner than May 20, 2028, so the embedding constant carries far more runway than the generation one. That row sits in the "Embeddings models" table, which the page shows without a click.

Dates for models Vertex AI already retired sit one level down, inside the expander labeled "The following table lists the retired models (click to expand)", so open the expander before you look for one of those. check_readme.py rejects a date that its transcription of these tables does not carry.

gemini-3.8-flash handles every prompt in this repository, including the ones that send a video or an image. That is a change from the earlier layout, where utils_vertex.py exported five model objects and the demos picked among three. Those objects, and the image generation and multimodal embedding models that sat beside them, are gone.

Vertex AI serves that model only on the global endpoint. A run against a live project sent one generation call per endpoint and got an answer from global and a 404 from us-central1, us-east5, southamerica-east1, us-east1, us-south1, and europe-southwest1. Every other Gemini 3.x flash identifier behaved the same way, and only the previous model generation, which this repository does not pin, answered regionally. The module docstring in utils_vertex.py dates that run. Five demos used to offer a region picker; every option it listed now fails, so the picker is gone and each of those pages states the endpoint instead.

Key dependencies

requirements.txt lists 15 direct dependencies. The demos depend most directly on these packages:

  • streamlit renders every demo page.
  • google-genai carries every Gemini call, in Vertex AI mode.
  • GitPython clones the repository that the repository inspection demo analyzes.
  • magika classifies the file types that the same demo reads.
  • pandas builds the token usage table that the same demo shows.

That list is neither complete nor fully pinned, so two installs from the same requirements.txt can give you different code.

apps/repo-inspection.py imports pandas to build that table. The manifest names pandas directly, so the demo survives a later streamlit release that drops it.

requirements.txt does not list nltk, nbconvert, nbformat, or PyPDF2. No source file imports them, and between them they carried most of the security advisories this repository reported, including four that no release fixes.

No source file imports tqdm either, but magika depends on it, so every install pulls it in. The tqdm>=4.66.3 line holds that dependency above the version the advisory covers, which a reinstall over an older environment otherwise keeps.

Eight of the 15 lines carry no version constraint, among them streamlit, google-genai, and pandas. Pin them before you depend on a build that has to keep working.

Set up your environment

  1. Clone the repository and change into it:

    git clone https://github.com/duboc/gemini-dev-demos.git
cd gemini-dev-demos

  2. Create and activate a virtual environment:

    python3 -m venv .venv
source .venv/bin/activate

  3. Install the dependencies:

    pip install -r requirements.txt

  4. Authenticate with application default credentials:

    gcloud auth application-default login

  5. Export the environment variables that the google-genai client reads:

    export GOOGLE_CLOUD_PROJECT="my-project"
export GOOGLE_CLOUD_LOCATION="global"
export GOOGLE_GENAI_USE_VERTEXAI=true

    Replace my-project with your Google Cloud project ID. Leave the location set to global. Vertex AI serves the model this repository pins only on the global endpoint, and a regional value such as us-central1 makes every call answer with a not-found error. The third variable points the client at Vertex AI instead of the Gemini Developer API, which takes an API key.

Environment variables

No file in this repository calls os.environ for any of the first three variables. The google-genai client reads them itself when get_client() constructs it, which is why a typo shows up as an authentication error from the SDK rather than as a None inside a demo.

Variable Set or read by Purpose
GOOGLE_CLOUD_PROJECT Dockerfile and cloudbuild.yaml set it, and the google-genai client reads it Google Cloud project that bills and serves the Gemini calls.
GOOGLE_CLOUD_LOCATION Dockerfile and cloudbuild.yaml set it, and the google-genai client reads it Vertex AI endpoint. Both files set it to global, the only endpoint that serves the pinned model. No demo overrides it.
GOOGLE_GENAI_USE_VERTEXAI Dockerfile and cloudbuild.yaml set it Set it to true to route the client through Vertex AI and application default credentials.
DEMO_ASSETS_BUCKET Dockerfile and setup.sh set it Names a Cloud Storage bucket for demo assets. No file in this repository reads the variable, so setting it changes nothing.
STREAMLIT_SERVER_ENABLE_STATIC_SERVING Dockerfile sets it Flag that lets Streamlit serve files from a static directory.

Run the demos locally

Start the Streamlit server from the repository root:

streamlit run home.py

Streamlit reads .streamlit/config.toml, which sets port 8080. Open http://localhost:8080 in your browser to use the demos.

home.py renders the shell for every demo. Use the Select a category list in the sidebar to pick a category, then click a demo button to load that page. If the app reports a state error, click Reset All in the sidebar to clear the Streamlit session state.

Every page renders on its own. A page that calls Gemini needs application default credentials and a project, so run gcloud auth application-default login and export the variables above before you press a generate button. For the identifiers those calls use, see Gemini model identifiers.

Available demos

Every demo lives in its own file under apps/ and loads inside the home.py shell. The first column gives the label on the sidebar button that opens the demo.

Sidebar button File Description
Repo Inspection apps/repo-inspection.py Clones a Git repository, classifies its files with Magika, generates an analysis of the codebase, and reports the tokens the run consumed.
Image to Code, Test and Deploy apps/code-to-image.py Turns an uploaded image of an application screen into a description, backend code, frontend code, deployment commands, test cases, and a Selenium script.
Cobol to Java apps/cobol-to-java.py Migrates a COBOL sample to Java in reviewable steps, then merges the steps into one program.
Selenium Automation apps/selenium-automation.py Generates a Selenium test script from a video of a user session.
Firebase Robo Script apps/firebase-testlab.py Generates a Firebase Test Lab Robo script from a video of an app walkthrough.
Appium Automation apps/appium-automation.py Generates an Appium script for mobile testing from a video of an app walkthrough.
UX Heuristic Analysis using Gemini AI apps/ux-heuristics-app.py Reviews an interface against usability heuristics and reports the findings.
UX Friction Log Generator apps/ux-frictionlog-app.py Produces a friction log from a recording of a user interaction.
Accessibility with Gemini apps/ux-accessibility.py Analyzes an interface against WCAG guidance, then turns the findings into accessibility user stories.
User Story to Code apps/generate-story B1AA -to-code-generic.py Generates a user story, breaks it into tasks, then generates code and a unit test implementation.
User Story to Data apps/generate-story-to-data-generic.py Generates a user story, breaks it into tasks, then generates a data warehouse model and a BigQuery implementation.
User Story to API apps/generate-story-to-api-generic.py Generates a user story, breaks it into tasks, then generates an OpenAPI specification and an Apigee implementation.
Dataform ELT Generation apps/dataform-gen.py Takes a schema that you enter and generates Dataform SQL and the matching Terraform definitions.

The three user story demos read seed prompts from data/. That directory holds one text file per industry and language, such as data/retail-en.txt, data/finance-pt.txt, and data/health-es.txt.

home.py defines six categories, and the sidebar shows the buttons for the category you select. The last category, Others, repeats the Dataform ELT generation demo, so the six categories list 14 entries for 13 demo files.

Repository layout

Path Contents
home.py Streamlit entry point: sidebar, category list, and demo loader.
apps/ One Python file per demo.
apps/firebase/ Support modules for the Firebase Robo script demo: apps/firebase/config.py, apps/firebase/generation.py, and apps/firebase/ui_components.py.
apps/firebase/prompts/ Markdown prompt templates for that demo.
utils_vertex.py Shared Gemini access: the google-genai client, the model identifiers, the safety settings, and the sendPrompt() helper.
utils_streamlit.py Helpers that reset Streamlit session state.
snippets/model-request.py Minimal example of calling sendPrompt().
data/ Seed text for the user story demos.
images/ Logo and animation that the interface displays.
.streamlit/config.toml Streamlit server, browser, and theme settings.
Dockerfile Container image definition.
cloudbuild.yaml Cloud Build pipeline that builds, pushes, and deploys the image.
setup.sh One-shot script that enables APIs, creates a bucket and an Artifact Registry repository, and deploys to Cloud Run from source.
check_readme.py Verification script for this README.
docs/CONTRIBUTING.md Contribution guide.
LICENSE Apache License 2.0 text.

Run the demos in a container

  1. Build the image:

    docker build -t gemini-dev-demos .

  2. Run the container, overriding the placeholder values that Dockerfile bakes in:

    docker run -p 8080:8080 \
  -e GOOGLE_CLOUD_PROJECT="${GOOGLE_CLOUD_PROJECT}" \
  -e GOOGLE_CLOUD_LOCATION="${GOOGLE_CLOUD_LOCATION}" \
  gemini-dev-demos

The image exposes port 8080 and starts streamlit run home.py. Open http://localhost:8080 to use the demos.

The container has no Google Cloud credentials of its own. Mount your application default credentials into the container or run it on a Google Cloud service that provides a service identity.

Deploy to Cloud Run

The repository offers two deployment paths. Both deploy the same application.

Deploy from source with setup.sh

setup.sh enables the required APIs, creates a Cloud Storage bucket, creates an Artifact Registry repository named dev-lifecycle, and deploys a Cloud Run service named dev-lifecycle from source with DEMO_ASSETS_BUCKET set to the new bucket. No demo reads that variable, so the bucket stays empty.

  1. Edit setup.sh and replace YOUR_PROJECT_ID and YOUR_REGION with your project ID and region.

  2. Run the script:

    bash setup.sh

Deploy with Cloud Build

  1. Create the Artifact Registry repository that cloudbuild.yaml expects:

    gcloud artifacts repositories create gemini-dev-demos --repository-format=docker --location=us-central1 --description="Gemini developer lifecycle demos"

  2. Submit the build:

    gcloud builds submit . --config=./cloudbuild.yaml --substitutions SHORT_SHA=1.0

cloudbuild.yaml installs the dependencies, builds the image, pushes it to Artifact Registry, and deploys it to Cloud Run with --allow-unauthenticated and with GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_LOCATION, and GOOGLE_GENAI_USE_VERTEXAI set on the service.

Override these substitution variables to change the target:

Substitution Default Purpose
_ARTIFACT_REGISTRY_REPO gemini-dev-demos Artifact Registry repository name.
_REPO_LOCATION us-central1 Artifact Registry location.
_SERVICE_NAME gemini-re-demos Cloud Run service name and image name.
_SERVICE_REGION us-central1 Cloud Run region.
_VERTEX_LOCATION global Vertex AI endpoint the deployed service calls.

_SERVICE_REGION and _VERTEX_LOCATION name two different things, and the pipeline used to pass one value for both. _SERVICE_REGION places the container; _VERTEX_LOCATION becomes GOOGLE_CLOUD_LOCATION on the service and picks the Vertex AI endpoint. A service that runs in us-central1 still calls the global endpoint, because that is the only endpoint serving the pinned model. Move the service wherever you like and leave the endpoint alone.

Cloud Build also requires SHORT_SHA, which it populates automatically for trigger-driven builds and which you pass yourself for manual builds.

The build step hardcodes the us-central1 Artifact Registry host, whereas the push and deploy steps use _REPO_LOCATION. If you change _REPO_LOCATION, update the build step in cloudbuild.yaml as well.

Verify this README

check_readme.py checks this README against the repository: the sections it must contain, referenced paths, demo coverage, sidebar labels, model identifiers, the Vertex AI endpoint, dated claims, environment variables, packages, deployment commands, the license statement, and a set of mechanical style rules. It needs no arguments, no network access, and no Google Cloud credentials.

python3 check_readme.py

The script prints one line per check and exits with a non-zero status if any check fails. Run it after you change README.md or rename anything it references.

Troubleshooting

  • Vertex AI returns a permission or API error. Confirm that the project named by GOOGLE_CLOUD_PROJECT has aiplatform.googleapis.com turned on, and that your credentials carry the Vertex AI User role.
  • The client reports missing credentials or no project. get_client() builds the google-genai client on the first call, not at import, so a missing variable surfaces when you press a generate button rather than when the page loads. Run gcloud auth application-default login, export GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_LOCATION, and GOOGLE_GENAI_USE_VERTEXAI, then restart the server.
  • A model call returns a not-found error. Two different causes produce the same error. Check the endpoint first: run echo $GOOGLE_CLOUD_LOCATION and set it to global if it names a region, because a region serves no Gemini 3.x model and changing which region you pick cannot help. If the endpoint is already global, the identifier is the suspect: Vertex AI serves gemini-3.8-flash for a short availability window, so confirm it against Gemini model identifiers.
  • A repository inspection run exhausts memory. apps/repo-inspection.py loads a whole repository into the prompt. Analyze a smaller repository, or run the app on a machine with more memory.
  • The container reports the wrong project. Dockerfile bakes in GOOGLE_CLOUD_PROJECT=my-demo-project-400313. Override it with -e GOOGLE_CLOUD_PROJECT when you run the container, or with --update-env-vars when you deploy.
  • Streamlit reports a session state error. Click Reset All in the sidebar, which clears every key in the session state.
  • Streamlit warns that it found no static folder. .streamlit/config.toml sets enableStaticServing, and the repository ships no static directory, so the server prints the warning on every start. Ignore it, or create the directory if you want to serve your own files from it.

Contribute

Contributions are welcome. Read the contribution guide for the Contributor License Agreement, the community guidelines, and the pull request process. Run python3 check_readme.py before you send a change that touches this README.

Security

  • Grant service accounts the smallest set of roles that the demos need.
  • Prefer gcloud auth application-default login or an attached service identity over downloaded service account keys.
  • Keep credential files out of the repository directory. .gitignore, .dockerignore, and .gcloudignore each exclude .env, and .gitignore also excludes *-credentials.json, but none of them excludes a service account key saved under another name, so such a key reaches your commits and your build context.
  • Remove --allow-unauthenticated from cloudbuild.yaml and setup.sh, or put the service behind Identity-Aware Proxy, before you expose a deployment beyond a demo audience.
  • Upload only material that you are allowed to send to Vertex AI. The demos forward your uploads, including source code and videos, to the model.

License

This project is licensed under the Apache License 2.0. For the full text, see the LICENSE file.

About

No description, website, or topics provided.

Resources

Readme
Apache-2.0 license

Contributing

Contributing
Activity

Stars

5 stars

Watchers

6 watching

Forks

0 forks
Report repository

Releases

Packages

Used by

Contributors

Languages
0