Tutorial: your first detection pipeline

Watch the whole flight first, then fly it yourself step by step. By the end, a detection pipeline is running live on your Axelera card — built by Agent Dev from one plain-English ask.

Full walkthrough, narrated

Build the pipeline step by step

You will not write YAML or pick a model in this tutorial. You describe the outcome, Agent Dev proposes a plan, and nothing executes until you approve it. The result is not just a window with boxes: every launch lands as a packaged folder with a rerunnable script, validation artifacts, and a browser viewer.

  1. Open Agent Dev in your Voyager workspace

    Open the desktop app, pick your Voyager workspace, and select Agent Dev. Only a compact green dot proves the selected provider is ready. Compact amber checking or degraded dots are diagnostic states, not proof of readiness; click one for details or a retry. If the control expands to Fix now, click it once and Wingman walks through the guided install, sign-in, or connection fix.

  2. Tell Agent Dev what you want

    Ask in plain words. Name the source you want — a USB camera, a video file — and where it should run.

    prompt

    Build a people-detection pipeline for my USB camera
and run it on the Axelera card.

  3. Approve the plan

    The agent proposes model, input, and output before anything runs. Nothing executes until you approve — permissions stay in your hands. Approve the plan and the first execution request.

  4. Watch it run on silicon

    The pipeline compiles, deploys, and starts on your card. Watch the run output for the proof that silicon, not CPU, is doing the work.

    output

    ✓ pipeline running on metis-0-1-0
✓ 15.6 FPS end-to-end · core 33°C

  5. Open the packaged result

    Every launch produces a package folder: RUN_HERE.sh reruns the pipeline any time, results/ holds the logs and validation files, and viewer/index.html is a browser dashboard that opens automatically after a validated run. A VALIDATED.txt file plus results/METIS_PROOF.json confirm the run really executed on the card.

    bash

    ./RUN_HERE.sh          # rerun the validated pipeline
ls results/            # logs, EVALUATION.json, METIS_PROOF.json
xdg-open viewer/index.html

Read the results viewer

View your results

The viewer shows end-to-end FPS separately from device FPS, plus latency and per-stage timings. Every metric carries a small info icon that names the artifact it was measured from — if a number cannot be traced to the run, the viewer refuses to show it. Video results play as extracted frames, so they render in any browser.

No card yet

Which hardware this covers

Ask, approve, verify

  • One ask, whole pipeline: Agent Dev turns a plain-English ask into a running Voyager pipeline — model choice, compilation, and execution included.
  • You stay in control: Every execution step waits for your approval first. You saw the plan before anything touched the card.
  • Proof, not vibes: The package folder, METIS_PROOF.json, and the browser viewer confirm the card did the work — and let you rerun it any time.

Continue to cameras & streams

Next rung: point this pipeline at real cameras and streams — Tutorial: run pipelines on cameras & streams.