inferctl

Project description

inferctl is a CLI-first sentiment scorer that grows into a full MLOps stack: package → API → train → track → version → deploy → schedule → monitor.

You start with a single command that scores text, then week by week add a Flask API and SDK, real Hugging Face train/eval, MLflow tracking and registry, DVC pipelines, Docker Compose deploy, tests and CI, orchestration, and finally monitoring plus rollback.

This README is the 8-week, Monday–Friday map. Each weekday is about 2.5 hours (~12.5 h/week). Friday is the exam: if the gate fails, do not start the next week. The spine is always inferctl — learn a slice, then put it in the tool the same day.

How to use

Install in editable mode, then call the CLI:

python -m venv .venv && source .venv/bin/activate
pip install -e .
inferctl --help

Core commands (built across the 8 weeks):

inferctl predict   --text "..." [--url|--model-id|--model-dir]
inferctl serve     --model-dir|--model-id [--host --port]
inferctl train     --data ... --out models/ [--epochs]
inferctl evaluate  --model-dir ... --split test
inferctl publish   --repo YOU/name
inferctl repro                  # dvc repro
inferctl register  --run-id ... # MLflow registry
inferctl rollback  --stage Production --version N
inferctl status

Quick start once Week 1 is done:

inferctl predict --text "this is great"

Work the roadmap top to bottom. Each week ends with a Friday exam — run those commands cold in a fresh terminal. If the gate fails, redo that week; do not skip ahead. Log results in notes/week-N.md.

Timeline

Week 1  CLI + package + predict
Week 2  Flask API + SDK + OpenAPI
Week 3  Real HF train / eval
Week 4  MLflow tracking + registry
Week 5  DVC data + pipelines
Week 6  Docker + Compose deploy
Week 7  Tests, CI, orchestration
Week 8  Monitor + rollback + demo

Optional 9–12  Staging/prod, batch, PEFT/Hub, canary

Horizon	Length	What you have
Spike	3.5 days	Toy CLI that predicts
This map	8 weeks	Train → track → version → deploy → schedule → monitor
Production-shaped	12 weeks	Staging/prod, drift, rollback, A/B

Target stack

git + inferctl CLI
        │
        ▼
   DVC pipeline          prepare → train → evaluate
        │
        ▼
   MLflow                runs, metrics, model registry
        │
        ▼
   Hugging Face Hub      checkpoints + model card
        │
        ▼
   Docker Compose        Flask scorer + MLflow UI
        │
        ▼
   Prefect / GitHub Actions    schedule + CI
        │
        ▼
   Prometheus / Grafana + smoke/rollback

How to use Friday every week

Run the exam commands cold (new terminal, activate venv, no leftover servers).
Fill this 5-line log in notes/week-N.md:

Gate: PASS / FAIL
What I can now do:
What I still fake:
Link/run_id/screenshot:
Start next week? yes / no (if no: Saturday redo, do not skip ahead)

If Friday fails, the next week’s Monday is a redo of the failed gate, not new material.

Week 1 — CLI on PATH + `predict`

Week outcome: pip install -e . && inferctl predict --text "this is great" works.

Day	Learn	From	Implement	How
Mon	Shell you will live in: paths, venv, `PATH`, pipes, exit codes	https://missing.csail.mit.edu/ · https://linuxcommand.org/tlcl.php (chs. 1–10)	Repo skeleton	`git init` if needed. Create `src/inferctl/`, `.gitignore`, venv. Confirm `python -c "import sys; print(sys.executable)"` is the venv.
Tue	CLI parsing + packaging + `console_scripts`	https://docs.python.org/3/howto/argparse.html · https://typer.tiangolo.com/ · https://packaging.python.org/tutorials/packaging-projects/ · https://packaging.python.org/specifications/entry-points/	Installable CLI with `--help`	`pyproject.toml` with `[project.scripts] inferctl = "inferctl.cli:app"`. Typer (or argparse) group: `predict`, `serve`, `train`, `publish` as stubs. `pip install -e .` then `inferctl --help`.
Wed	What a Transformer pipeline is: tokenizer → model → postprocess	https://huggingface.co/docs/transformers/en/quicktour · https://huggingface.co/learn/llm-course/en/chapter1/1 · https://huggingface.co/learn/llm-course/en/chapter2/1	`src/inferctl/predict.py`	Load `distilbert-base-uncased-finetuned-sst-2-english` via `pipeline("sentiment-analysis")`. Function `predict_text(text, model_id) -> dict`. No Flask yet.
Thu	Logging, exit codes, not `print`	https://docs.python.org/3/library/logging.html · https://docs.python.org/3/library/pathlib.html	Wire `inferctl predict`	`--text`, `--model-id`. JSON to stdout. `logging` to stderr. Exit `2` on bad args, `1` on model/runtime error, `0` on success.
Fri	Exam — see below	—	Freeze Week 1	Tag `week-1` if the gate passes.

Friday exam (Week 1)

python -m venv .venv && source .venv/bin/activate
pip install -e .
inferctl --help
inferctl predict --text "this is great"
inferctl predict          # must fail with exit 2

Pass	Fail
`inferctl` is on PATH after editable install	Only works as `python src/...`
Happy path prints JSON with `label` + `score`	Raw Python traceback to the user
Missing `--text` exits `2`	Exits `0` or `1` on usage error
No `print()` in library code	Debug prints mixed into JSON

Week 2 — Flask API + SDK + OpenAPI

Week outcome: CLI in-process and curl return the same JSON.

Day	Learn	From	Implement	How
Mon	HTTP/REST: methods, status codes, JSON resources, what an API is vs an SDK	https://restfulapi.net/ · https://developer.mozilla.org/en-US/docs/Web/HTTP · https://learn.microsoft.com/en-us/azure/architecture/best-practices/api-design · https://academy.postman.com/path/api-beginner	Write the contract first	`openapi.yaml`: `GET /health`, `POST /predict` with request/response schemas. Open in https://editor.swagger.io/ — it must validate.
Tue	Flask app factory, blueprints, project layout	https://flask.palletsprojects.com/en/stable/tutorial/ · https://flask.palletsprojects.com/en/stable/tutorial/factory/ · https://flask.palletsprojects.com/en/stable/tutorial/views/ · https://flask.palletsprojects.com/en/stable/quickstart/	`src/inferctl/app.py`	`create_app()`. Blueprint with `/health` only. `flask --app inferctl.app:create_app run` returns `{"status":"ok"}`.
Wed	Mapping HTTP ↔ your predict function	Same Flask tutorial + MDN status codes	`POST /predict` + `inferctl serve`	Body `{"text":"..."}`. 400 on missing text, 200 + same dict as CLI. `inferctl serve --host 127.0.0.1 --port 8000` starts the factory.
Thu	SDK design: retries, timeouts, typed errors	https://newsletter.pragmaticengineer.com/p/building-great-sdks · https://auth0.com/blog/guiding-principles-for-building-sdks/	`src/inferctl/client.py`	`Client(base_url)` with `health()` and `predict(text)`. Timeouts. Retry once on 503. `inferctl predict --url http://127.0.0.1:8000 --text "..."`.
Fri	Exam	—	Contract freeze	Commit `openapi.yaml` + client.

Friday exam (Week 2)

inferctl serve --port 8000 &
curl -s http://127.0.0.1:8000/health
curl -s -X POST http://127.0.0.1:8000/predict -H 'Content-Type: application/json' -d '{"text":"this is great"}'
inferctl predict --text "this is great" > /tmp/cli.json
inferctl predict --url http://127.0.0.1:8000 --text "this is great" > /tmp/http.json
diff /tmp/cli.json /tmp/http.json
curl -s -o /dev/null -w '%{http_code}\n' -X POST http://127.0.0.1:8000/predict -d '{}'

Pass	Fail
`diff` is empty (same `label`/`score`/`model_id`)	CLI and HTTP disagree
Bad body → 400, not 500	Unhandled exception page
OpenAPI file opens in Swagger Editor	Spec is comments in README only
Client timeout/retry exist as real code	`requests.get` with no timeout

Week 3 — Real HF train / eval

Week outcome: a fine-tune beats the off-the-shelf baseline on a held-out slice.

Day	Learn	From	Implement	How
Mon	Models, tokenizers, datasets, Hub	https://huggingface.co/learn/llm-course/en/chapter1/1 through https://huggingface.co/learn/llm-course/en/chapter2/6 · https://huggingface.co/docs/datasets	`src/inferctl/data.py`	Load a small SST-2 slice (e.g. 1k/200 train/val) via `datasets`. Tokenize. Save processed arrows or a local cache path.
Tue	Fine-tuning with `Trainer`	https://huggingface.co/learn/llm-course/en/chapter3/1 · https://huggingface.co/docs/transformers/en/training	Notebook or script first, then port	Fine-tune `distilbert-base-uncased` for 1 epoch on the slice. Confirm `eval_accuracy` is logged. Do not wrap CLI until this runs.
Wed	Turning a script into a CLI command + reproducibility	https://huggingface.co/docs/accelerate · https://docs.python.org/3/library/logging.html	`inferctl train`	Flags: `--data`, `--out`, `--epochs`, `--model-id`, `--seed`. Write `models/.../config` + weights + `metrics.json`.
Thu	Evaluation vs prediction; config files	HF course ch. 3 (evaluation) · https://huggingface.co/docs/datasets · TOML via stdlib	`inferctl evaluate` + `train.toml`	`evaluate --model-dir --split test` writes metrics. Train can take `--config train.toml`. Seed is set (`random`, `numpy`, `torch`).
Fri	Exam	—	Champion vs baseline	Compare pipeline baseline vs your `--model-dir`.

Friday exam (Week 3)

inferctl train --data sst2 --out models/week3 --epochs 1 --seed 42
inferctl evaluate --model-dir models/week3 --split validation
inferctl predict --model-dir models/week3 --text "terrible movie"
# baseline (pretrained SST-2 head) vs your run — record both accuracies in notes/week-3.md

Pass	Fail
`models/week3` exists and `predict --model-dir` uses your weights	Still silently hitting the Hub pretrained model
`metrics.json` has accuracy (or F1)	Only loss printed in the terminal
Seed flag is wired	Cannot rerun the same experiment
Your eval ≥ baseline or you write why it lost (data too small is acceptable if honest)	“It trained” with no number

Week 4 — MLflow tracking + registry

Week outcome: every train is a run; a champion is registered and servable.

Day	Learn	From	Implement	How
Mon	Why experiment tracking exists; runs vs models vs registry	https://mlflow.org/docs/latest/index.html (Get Started + Tracking) · Zoomcamp Module 2: https://github.com/DataTalksClub/mlops-zoomcamp/tree/main/02-experiment-tracking · https://madewithml.com/courses/mlops/	Local MLflow	`pip install mlflow` then `mlflow ui`. Log a dummy run by hand (`mlflow.start_run`, log param/metric).
Tue	Logging from training code	MLflow Python API (same docs) · Zoomcamp M2 homework as reference	Instrument `inferctl train`	On each train: log params (epochs, lr, seed, model_id), metrics from `Trainer`, artifacts (`metrics.json`, model dir). `MLFLOW_TRACKING_URI` env.
Wed	Model registry + stages	https://www.mlflow.org/docs/latest/ml/model-registry/	`inferctl register`	`--run-id` `--name inferctl-sst2` `--stage Staging`. Use `mlflow.register_model` + transition stage.
Thu	Serving a registry URI	MLflow Models / pyfunc docs · Week 2 Flask app	`serve` loads `models:/inferctl-sst2/Staging`	`--model-uri models:/inferctl-sst2/Staging`. `/health` reports the URI + version.
Fri	Exam	—	Two runs, one champion	`notes/week-4.md` names the winning `run_id`.

Friday exam (Week 4)

inferctl train --data sst2 --out models/w4a --epochs 1 --seed 1
inferctl train --data sst2 --out models/w4b --epochs 2 --seed 2
mlflow ui   # visually confirm 2 runs
inferctl register --run-id <winner> --name inferctl-sst2 --stage Staging
inferctl serve --model-uri models:/inferctl-sst2/Staging --port 8000
curl -s http://127.0.0.1:8000/health

Pass	Fail
UI shows two runs with params and metrics	Only local folders, nothing in MLflow
Registry has `inferctl-sst2` in Staging	Model copied by hand into Flask
`/health` reports registry version	You cannot tell which run is live
Train without MLflow still works if URI unset (or fails loudly)	Hard crash with an opaque stack trace

Week 5 — DVC data + pipelines

Week outcome: dvc repro is how you train, not a notebook.

Day	Learn	From	Implement	How
Mon	Why data cannot live in git; `.dvc` files, remotes	https://dvc.org/doc/start · https://dvc.org/doc/start/data-management	Init + track raw data	`dvc init`. Put a small raw SST-2 export under `data/raw/`. `dvc add data/raw`. Commit the `.dvc` file, not the bytes.
Tue	Pipeline stages, deps, outs	https://dvc.org/doc/start/data-pipelines	`prepare` stage	`dvc.yaml` stage `prepare` runs `inferctl` (or a small module) and writes `data/processed`. `dvc repro` builds it.
Wed	Chaining train/eval as stages	Same DVC pipelines page · Zoomcamp if the current cohort covers DVC	`train` + `evaluate` stages	`train` deps on processed data; outs `models/dvc` + metrics. `evaluate` deps on the model. Commands must call `inferctl`, not hidden scripts.
Thu	Metrics, plots, remotes	https://dvc.org/doc/user-guide/experiment-management/comparing-experiments · DVC remotes in the Start guide	`dvc metrics` + local remote	Declare metrics in `dvc.yaml`. `dvc remote add -d local .dvc-remote` (or `/tmp/dvc-remote`). `dvc push` / `dvc pull`.
Fri	Exam	—	`inferctl repro`	Thin wrapper around `dvc repro`.

Friday exam (Week 5)

dvc status
dvc repro
# change one processed-data input or a train hyperparam in params.yaml
dvc repro          # must skip unchanged stages, rerun downstream
dvc metrics show
dvc push
inferctl repro

Pass	Fail
`data/raw` is not in git history	Dataset committed to git
Second `dvc repro` after a param change retrains and skips `prepare` if data unchanged	Everything always reruns, or nothing does
`dvc metrics show` prints accuracy	Metrics only in MLflow or a screenshot
Fresh `dvc pull` after deleting `data/` restores files	Pipeline only works on your laptop’s cache

Week 6 — Docker + Compose

Week outcome: a stranger can docker compose up and hit /predict.

Day	Learn	From	Implement	How
Mon	Images vs containers, layers, `WORKDIR`, ports	https://docs.docker.com/get-started/	Run someone else’s image	`docker run --rm -p 8000:8000` a tiny hello image so Docker itself is not the blocker.
Tue	Writing a service Dockerfile	https://docs.docker.com/get-started/workshop/02_our_app/ · https://flask.palletsprojects.com/en/stable/tutorial/deploy/	`Dockerfile` for serve	Install from `pyproject.toml`. `CMD` = `inferctl serve --host 0.0.0.0 --port 8000`. `.dockerignore`. Non-root user.
Wed	Compose services + volumes	https://docs.docker.com/compose/	`compose.yaml`	Services: `api` (your image), `mlflow` (`mlflow ui` or server), named volume for artifacts. `docker compose up --build`.
Thu	Training in a container + version labels	Docker docs on `docker compose run`	`compose run train`	One-shot train service writes into the shared volume. Label image with git sha / model version. `/health` returns that version.
Fri	Exam	—	Clean-room boot	Remove local venv from the test path.

Friday exam (Week 6)

docker compose down -v
docker compose up --build -d
curl -sf http://127.0.0.1:8000/health
inferctl predict --url http://127.0.0.1:8000 --text "this is great"
# from a second terminal: compose logs show no traceback

Pass	Fail
Predict works without `pip install -e .` on the host	Container imports host site-packages
`/health` includes model/image version	“ok” with no identity
Image rebuild is deterministic enough to document	“It works on my Docker Desktop” only
`api` can read the model volume	Model baked in with no way to swap

Week 7 — Tests, CI, orchestration

Week outcome: a breaking PR cannot merge; training can be scheduled.

Day	Learn	From	Implement	How
Mon	Testing CLIs and Flask apps	https://docs.pytest.org/en/stable/how-to/ · https://flask.palletsprojects.com/en/stable/tutorial/tests/ · https://learn.scientific-python.org/development/	`tests/`	`pytest`: parse CLI help; `POST /predict` via Flask test client; one golden JSON fixture (tiny fake/stubbed pipeline if needed so CI is CPU-light).
Tue	CI as a gate	https://docs.github.com/en/actions/get-started/quickstart · Zoomcamp CI bits	`.github/workflows/ci.yml`	On PR: lint (`ruff`), `pytest`, optional OpenAPI parse. No GPU. Fixture data only.
Wed	Orchestration vs cron vs CI	https://docs.prefect.io/ · Zoomcamp orchestration module in https://github.com/DataTalksClub/mlops-zoomcamp	Prefect hello	Install Prefect. A flow that prints and calls `inferctl evaluate` on an existing model. Run once locally.
Thu	Promotion policy in a flow	Prefect docs: tasks, retries · Week 4 registry	Flow: `repro → read metrics → register if acc > threshold`	Threshold in config. If below, flow succeeds but does not promote (do not fail the world on a bad model).
Fri	Exam	—	Red-green CI + one flow run	Break a test on a branch and watch CI fail.

Friday exam (Week 7)

pytest -q
# push a branch that deletes an assertion or breaks /health
# CI must go red
# run the promote-or-skip Prefect flow once

Pass	Fail
`pytest` green on a clean tree	Only manual clicking
Intentional break → CI red	CI is `echo hello`
Flow registers only when metric beats threshold	Always registers, or always crashes
CI does not download a 400MB model every PR	Jobs timeout / burn the free minutes

Week 8 — Monitor, rollback, demo

Week outcome: you can promote, serve, watch, and undo. This is the finish line.

Day	Learn	From	Implement	How
Mon	SLIs for a model service: latency, errors, throughput	https://huyenchip.com/mlops/ · Zoomcamp monitoring module · https://prometheus.io/docs/instrumenting/exposition_formats/	`GET /metrics`	Counters: requests, errors. Histogram: latency. Use `prometheus_flask_exporter` or a 30-line manual exporter.
Tue	Dashboards	Grafana first-dashboard docs or a static HTML page hitting `/metrics`	One dashboard	Three panels: QPS, p95, error rate. Docker Compose service optional.
Wed	Prediction logs + cheap drift	Chip Huyen monitoring notes · https://madewithml.com/courses/mlops/	Log each request	Append JSONL: timestamp, text length, predicted label, score. On-demand job: if mean length or label mix shifts past a threshold, print `DRIFT`.
Thu	Rollback as a first-class op	MLflow registry stage transitions	`inferctl rollback` + `inferctl status`	`rollback --name inferctl-sst2 --stage Production --version N`. `status` prints live URI, registry version, last DVC metric, `/health`.
Fri	Exam — full stack demo	—	One script, ~4 minutes	Record commands + screenshot of MLflow + dashboard in `notes/week-8.md`.

Friday exam (Week 8) — the only exam that matters

git pull && dvc pull
inferctl repro
inferctl register --run-id <latest> --name inferctl-sst2 --stage Production
docker compose up -d --build
inferctl predict --url http://127.0.0.1:8000 --text "this is great"
curl -s http://127.0.0.1:8000/metrics | head
inferctl rollback --name inferctl-sst2 --stage Production --version 1
inferctl status

Pass	Fail
Script runs without hand-editing files mid-way	“Then I opened a notebook…”
`/metrics` moves when you send 20 requests	Empty endpoint
Rollback changes what `/health` reports	Rollback is `git checkout` of weights
README / notes can be followed in a new venv/compose	Works only in the original shell session

What not to do when

Temptation	Wait until
Kubeflow / K8s	Week 12, if ever
Feature store (Feast)	After week 8, only if you have real features
W&B instead of MLflow	Fine, but pick one tracker
Full Flask Mega-Tutorial auth/DB	Never required for this scorer
Giant LLM training	Never on this roadmap; use a classification / small LM
Perfect SDK generation	Week 2 handwritten client is enough

Weekly checkpoints

End of	Must be true
W1	`pip install -e . && inferctl predict --text "ok"`
W2	CLI and `curl` return the same JSON
W3	Fine-tuned weights beat the baseline on a held-out slice (or you document why not)
W4	Champion is in the MLflow registry
W5	`dvc repro` is how you train, not a notebook
W6	Stranger can `docker compose up` and hit `/predict`
W7	A PR that breaks tests cannot merge
W8	You can promote, serve, watch, and roll back

Optional weeks 9–12

Only after Week 8 is boringly reliable.

Week	Mon learn	Tue–Thu implement	Friday exam
9 Staging vs Production	https://huyenchip.com/mlops/ + MLflow aliases	Two Compose files / two registry stages; human approval before Production	Cannot promote to prod from laptop without an explicit `--yes` + stage check
10 Batch vs online	HF Datasets + your client	`inferctl batch --input data.csv --output preds.csv`	1k-row CSV scored offline; numbers match online `/predict` on 5 sampled rows
11 PEFT/LoRA + Hub registry	https://huggingface.co/docs/peft · https://huggingface.co/docs/huggingface_hub/guides/cli · https://huggingface.co/docs/trl	LoRA train on a slightly larger model; `inferctl publish --repo YOU/inferctl-lora`	`predict --model-id YOU/inferctl-lora` works on a clean machine after `huggingface-cli download`
12 Canary / optional K8s	Huyen serving + Docker networking	90/10 traffic split between two `--model-uri`s or a single k8s Deployment	Kill canary; error rate stays flat; rollback still works

How this maps to the five learning phases

Phase	Topics	Weeks	Outcome
1. Foundation	CLI + shell automation	1	You live in the terminal; `inferctl` is on PATH
2. Interfaces	APIs → SDKs	2	You consume and wrap your own service
3. Python CLIs	argparse → automation → packaging	1–2	You ship `inferctl`
4. Web	Flask	2 + 6	You serve `/predict` locally and in Docker
5. ML	HF Transformers → full MLOps	3–8 (9–12 optional)	You train, track, version, deploy, schedule, monitor

Resource index

Topic	URL
Missing Semester	https://missing.csail.mit.edu/
Linux Command Line (free PDF)	https://linuxcommand.org/tlcl.php
argparse tutorial	https://docs.python.org/3/howto/argparse.html
Typer	https://typer.tiangolo.com/
PyPA packaging	https://packaging.python.org/tutorials/packaging-projects/
Entry points	https://packaging.python.org/specifications/entry-points/
logging	https://docs.python.org/3/library/logging.html
pathlib	https://docs.python.org/3/library/pathlib.html
REST	https://restfulapi.net/
MDN HTTP	https://developer.mozilla.org/en-US/docs/Web/HTTP
Microsoft API design	https://learn.microsoft.com/en-us/azure/architecture/best-practices/api-design
Postman Academy	https://academy.postman.com/path/api-beginner
OpenAPI	https://spec.openapis.org/oas/latest.html
Swagger Editor	https://editor.swagger.io/
Flask tutorial	https://flask.palletsprojects.com/en/stable/tutorial/
Flask factory	https://flask.palletsprojects.com/en/stable/tutorial/factory/
Flask views	https://flask.palletsprojects.com/en/stable/tutorial/views/
Flask tests	https://flask.palletsprojects.com/en/stable/tutorial/tests/
Flask deploy	https://flask.palletsprojects.com/en/stable/tutorial/deploy/
Flask quickstart	https://flask.palletsprojects.com/en/stable/quickstart/
Building great SDKs	https://newsletter.pragmaticengineer.com/p/building-great-sdks
Auth0 SDK principles	https://auth0.com/blog/guiding-principles-for-building-sdks/
HF LLM Course ch.1	https://huggingface.co/learn/llm-course/en/chapter1/1
HF LLM Course ch.2	https://huggingface.co/learn/llm-course/en/chapter2/1
HF LLM Course ch.3	https://huggingface.co/learn/llm-course/en/chapter3/1
Transformers quicktour	https://huggingface.co/docs/transformers/en/quicktour
Transformers training	https://huggingface.co/docs/transformers/en/training
Datasets	https://huggingface.co/docs/datasets
Accelerate	https://huggingface.co/docs/accelerate
PEFT	https://huggingface.co/docs/peft
TRL	https://huggingface.co/docs/trl
huggingface-cli	https://huggingface.co/docs/huggingface_hub/guides/cli
MLflow	https://mlflow.org/docs/latest/index.html
MLflow Model Registry	https://www.mlflow.org/docs/latest/ml/model-registry/
MLOps Zoomcamp	https://github.com/DataTalksClub/mlops-zoomcamp
Zoomcamp M2	https://github.com/DataTalksClub/mlops-zoomcamp/tree/main/02-experiment-tracking
Made With ML	https://madewithml.com/courses/mlops/
Chip Huyen MLOps	https://huyenchip.com/mlops/
DVC start	https://dvc.org/doc/start
DVC data	https://dvc.org/doc/start/data-management
DVC pipelines	https://dvc.org/doc/start/data-pipelines
Docker get started	https://docs.docker.com/get-started/
Docker Compose	https://docs.docker.com/compose/
pytest	https://docs.pytest.org/en/stable/how-to/
GitHub Actions quickstart	https://docs.github.com/en/actions/get-started/quickstart
Scientific Python dev	https://learn.scientific-python.org/development/
Prefect	https://docs.prefect.io/
Prometheus exposition format	https://prometheus.io/docs/instrumenting/exposition_formats/

Weekly notes

Copy this into notes/week-N.md every Friday:

Gate: PASS / FAIL
What I can now do:
What I still fake:
Link/run_id/screenshot:
Start next week? yes / no

Name		Name	Last commit message	Last commit date
README.md		README.md

Navigation Menu

Repository navigation

About

Repository files navigation

inferctl

Project description

How to use

Timeline

Target stack

How to use Friday every week

Week 1 — CLI on PATH + `predict`

Friday exam (Week 1)

Week 2 — Flask API + SDK + OpenAPI

Friday exam (Week 2)

Week 3 — Real HF train / eval

Friday exam (Week 3)

Week 4 — MLflow tracking + registry

Friday exam (Week 4)

Week 5 — DVC data + pipelines

Friday exam (Week 5)

Week 6 — Docker + Compose

Friday exam (Week 6)

Week 7 — Tests, CI, orchestration

Friday exam (Week 7)

Week 8 — Monitor, rollback, demo

Friday exam (Week 8) — the only exam that matters

What not to do when

Weekly checkpoints

Optional weeks 9–12

How this maps to the five learning phases

Resource index

Weekly notes

About

Releases

Packages

Contributors

Latest commit

History

Folders and files

Repository files navigation

inferctl

Project description

How to use

Timeline

Target stack

How to use Friday every week

Week 1 — CLI on PATH + predict

Friday exam (Week 1)

Week 2 — Flask API + SDK + OpenAPI

Friday exam (Week 2)

Week 3 — Real HF train / eval

Friday exam (Week 3)

Week 4 — MLflow tracking + registry

Friday exam (Week 4)

Week 5 — DVC data + pipelines

Friday exam (Week 5)

Week 6 — Docker + Compose

Friday exam (Week 6)

Week 7 — Tests, CI, orchestration

Friday exam (Week 7)

Week 8 — Monitor, rollback, demo

Friday exam (Week 8) — the only exam that matters

What not to do when

Weekly checkpoints

Optional weeks 9–12

How this maps to the five learning phases

Resource index

Weekly notes

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Week 1 — CLI on PATH + `predict`