The jobs a general purpose model keeps getting almost right.
Prompting and retrieval get you to a demo. Narrow, repetitive, high volume tasks are where a small fine tuned model wins on accuracy, latency and cost at the same time. These are the shapes that fit.
Structured extraction from messy documents
Invoices, purchase orders, bills of lading and remittance advice arrive as scans with inconsistent layouts. A tuned model returns the same strict JSON every time instead of drifting between formats.
- MODEL
- Llama 3.1 8B or Qwen2.5 VL
- DATA
- Document scans plus labels
- OUTPUT
- Strict JSON schema
Defect detection from line photos
Surface scratches, misaligned seals and missing components are hard to spot consistently, and a general detector does not know your parts. Train one on your own labelled line photos and download the detector weights to run wherever your inspection stack already lives.
- MODEL
- YOLO
- DATA
- Kaggle or your own labelled images
- OUTPUT
- Downloadable detector weights
Ticket triage in your own taxonomy
Generic classifiers do not know your product areas, severity ladder or routing rules. A small tuned model learns your label set from historical tickets and runs cheaply at high volume.
- MODEL
- Small instruct model with LoRA
- DATA
- Exported ticket history
- OUTPUT
- Label, severity, queue
Shelf, pallet and inventory audits
Counting facings, spotting out of stock positions and reading pallet labels from phone photos taken in bad light. Off the shelf detectors miss your SKUs. A tuned one does not.
- MODEL
- YOLO detection
- DATA
- Store photos plus labels
- OUTPUT
- Counts and positions
Domain forms and clause parsing
Intake forms, referral letters and contract clauses use vocabulary that general models paraphrase away. Fine tuning holds the wording and the structure that downstream systems depend on.
- MODEL
- Llama or Qwen with LoRA
- DATA
- Redacted internal corpus
- OUTPUT
- Field level extraction
Domain specific coding and config copilots
Internal DSLs, infrastructure templates and legacy frameworks are underrepresented in pretraining. A tuned adapter on your own repositories writes code your team actually merges.
- MODEL
- Code base model with LoRA
- DATA
- Repo snapshots
- OUTPUT
- Completions and diffs
When to fine tune, and when to keep prompting.
Fine tuning is not always the answer. Here is the honest split, so you do not burn a weekend on a run you did not need.
- The same narrow task runs thousands of times a day
- You need strict, machine parseable output every single time
- Latency or unit cost makes a frontier API impractical
- The domain vocabulary is not in general pretraining data
- The work is visual and has to run on device or at the edge
- Data cannot leave your infrastructure
- Open ended reasoning across constantly changing topics
- Answers that must reflect documents updated every hour
- A one off exploration you will run a handful of times
- You have no examples and no raw material you could label
Same agent, different reason to reach for it.
Skip the MLOps detour
You already know how to ship services. Tensovy removes the part of the job that requires CUDA debugging, config archaeology and cluster babysitting, and gives you back a weights file and a registered model.
A first model in a week, not a quarter
Get a differentiated model into the product without the infrastructure detour that normally comes first. Run it on Tensovy managed infrastructure or connect your own cloud account, see the hourly cost before anything starts, and keep the weights either way.
Delete the glue work
You could write the pipeline yourself. You have written it before. Hand off provisioning, smoke testing, training and export, and still review every dataset and config before a run starts.
Prompts you can paste on day one.
- > fine tune a model that turns support emails into a JSON object with intent, product area and urgency
- > train a detector that finds cracked solder joints in these 400 board photos
- > build a model that rewrites our legacy config files into the new schema
- > extract vendor, line items and totals from these scanned invoices as strict JSON
- > classify incoming resumes against our eight internal role families
- > train a vision model to read gauge readings from plant inspection photos
Have a task in mind already?
Bring it with you. Early access is open, and the task you describe on the form is what your first run starts from.