08 / Profiles and models

Same agent, different model lanes.

Profiles change the model and auth path. The JSON agent and Cargo-AI tool stay fixed.

List

Check profiles

Start with cargo ai profile list so the model choice is explicit.

Run

Default GPT

Run the five photos one at a time into the GPT comparison CSV.

Compare

Gemini profile

Run the same five photos again with a Gemini profile and a separate CSV.

default GPT profilemanual runs

cargo ai profile list

cargo ai run --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/01.jpg \
  --input-text "Image name: 01.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gpt.csv

cargo ai run --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/02.jpg \
  --input-text "Image name: 02.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gpt.csv

cargo ai run --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/03.jpg \
  --input-text "Image name: 03.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gpt.csv

cargo ai run --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/04.jpg \
  --input-text "Image name: 04.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gpt.csv

cargo ai run --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/05.jpg \
  --input-text "Image name: 05.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gpt.csv

Gemini profilereplace profile name

cargo ai run --profile YOUR_GEMINI_PROFILE --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/01.jpg \
  --input-text "Image name: 01.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gemini.csv

cargo ai run --profile YOUR_GEMINI_PROFILE --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/02.jpg \
  --input-text "Image name: 02.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gemini.csv

cargo ai run --profile YOUR_GEMINI_PROFILE --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/03.jpg \
  --input-text "Image name: 03.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gemini.csv

cargo ai run --profile YOUR_GEMINI_PROFILE --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/04.jpg \
  --input-text "Image name: 04.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gemini.csv

cargo ai run --profile YOUR_GEMINI_PROFILE --config animal-patrol.json \
  --input-mode append \
  --input-image inputs/model-comparison-photos/05.jpg \
  --input-text "Image name: 05.jpg" \
  --run-var output_csv_path=outputs/animal-patrol-manual-gemini.csv

Output results

Five images, two model lanes, same CSV shape.

The same agent and Cargo-AI tool wrote both files. That makes the model comparison concrete instead of anecdotal.

Default profile

my_open_ai · gpt-5.5

Five runs completed in about 20.5 seconds total.

Local profile

gemma_4 · gemma4:e4b

Five runs completed in about 4 minutes total.

outputs/animal-patrol-manual-gpt.csvgpt-5.5

image_name	animal	count	confidence	garden_risk	notes
01.jpg	none	0	0	none	No animal visible in the garden scene.
02.jpg	none	0	0	none	No animal visible in the garden scene.
03.jpg	budgerigars (parakeets)	2	0.98	low	Two budgerigars perched on wood; small seed-eating birds with minimal garden impact.
04.jpg	squirrel	1	0.98	medium	One squirrel beside flowering plants; may dig in beds or nibble bulbs, seeds, and young growth.
05.jpg	deer	1	0.98	high	Spotted deer visible; deer commonly browse vegetables, flowers, and young shrubs.

outputs/animal-patrol-manual-gemini.csvgemma4:e4b

image_name	animal	count	confidence	garden_risk	notes
01.jpg	none	0	0	none	No animals visible. The image shows a garden setting with a clear sky and distant landscape.
02.jpg	none	0	0	low	The model shifted to path and overgrowth risk instead of animal risk.
03.jpg	parrot	1	0.9	low	Notes describe two birds, but the count field says one.
04.jpg	squirrel	1	0.95	medium	The squirrel is visible near flowers and pots; it may dig or consume seeds and berries.
05.jpg	deer	1	0.95	low	The deer is near a feeding station and may be attracted to the garden.

Agreement

Both runs found the main subjects.

Both outputs recognized no animal in the first photo, birds in the third, a squirrel in the fourth, and a deer in the fifth.

Difference

GPT stayed closer to the task boundary.

GPT kept no-animal images at risk `none`, identified two budgerigars, and rated deer as high garden risk.

Drift

The local run exposed useful edge cases.

The local model assigned path risk on a no-animal image and described two birds while recording count `1`.

Point

The tool makes comparison easy.

Different models can disagree, but the Cargo-AI tool preserves one stable CSV contract for inspection.