Specialized AI
Specialized AI covers the narrow jobs beside the big chat models: reading a receipt, translating a paragraph, speaking text aloud, running a small model on a phone or a tiny computer, and checking whether a code assistant fixes a bug. The steps below are checks you can run yourself, plus the public benchmarks that already publish a method.
Beginner
Read a receipt, translate a note, hear it aloud
In a photo of a receipt, the text is only pixels. Optical character recognition (OCR) is the step that turns that printed ink into letters and numbers. A chat model can then fill the fields you actually want: seller, date, total, and tax. Reading every word still leaves room to drop the tax into the wrong box.
Use a receipt you are allowed to share. A real slip can show a name, a card fragment, or a phone number. Cover those before you upload the photo to a chat in the phone browser. Ask for a short table with seller, date, currency, total, and tax, and tell the model to write "unreadable" when the print is smudged. Then check the total against the paper. Swedish receipts often use a comma for decimals and the word kronor or the abbreviation kr. If 149,00 comes back as 14900, count that as a miss.
You can do two other jobs in the same browser. Read-aloud uses the Web Speech API: the page sends a sentence to the device's speech service, and the speakers play it. You can pick a voice, including a Swedish one when the phone has it. Translation can be a paste into the same chat. Keep one rule: names, numbers, dates, and units stay as printed. Desktop Chrome can also translate with the built-in Translator API. That API downloads a language pack onto the computer, and Swedish is on the published list. Chrome's own guide says this API runs on desktop Chrome, so on a phone stay with the chat page.
Read one receipt without guessing
You are looking at a photo of a receipt or invoice that I am allowed to share. Extract only what is printed. Return a markdown table with columns: field, value, words you can see in the image. Fields: seller, date, currency, total, tax or VAT, payment method. If a field is missing, blurry, or cut off, write unreadable. Do not calculate a missing tax. Do not invent a seller from a logo when the name is not printed. Keep the original decimal mark (comma or dot). After the table, list any personal name, card fragment, phone number, or national ID number so I can cover it before sharing the image again.
Intermediate
Score one niche tool on a fixed set
Judge a niche tool by repeating the same inputs and writing down every miss. Ten items are enough to see a pattern. Keep an answer key beside the files: the true date, total, and tax, or the words a person actually said.
Score a receipt or invoice on two lines. The first line is the text read off the page (the OCR). The second is the fields: date, seller, total, tax, and currency. A full transcript with a wrong total is a failed extraction. For translation, keep the source and a careful human version, and mark a miss when a name, number, date, or unit changes. For speech, compare the transcript with the clip. The Open ASR Leaderboard reports word error rate, the share of words that differ from a reference transcript, and inverse real-time factor, a speed score where a higher number means more audio transcribed per second of work. Its tracks cover English, long recordings, and multilingual short clips. A Swedish clip you record still matters, because a strong average can hide one weak language.
For a code assistant, give one real bug, the broken behavior, and what fixed means. SWE-bench does this with real GitHub issues: the model must submit a patch, and the board scores the percent resolved. The site lists 2,294 instances in the full set and 500 in the human-checked Verified subset. Your trial can be one bug and a test you run. Record the date, the tool, the device, and each miss.
Score two translations on the facts
Compare Translation A and Translation B of the source below. Do not reward style. Check only these four points: names unchanged, numbers unchanged, dates unchanged, units unchanged. Then list meaning misses in plain words, each tied to a phrase in the source. End with a table: check, A, B, pass or miss. Source language: [language] Target language: [language] Source: [paste source] Translation A: [paste A] Translation B: [paste B]
Expert
Small models, browser runtimes, strict fields
A parameter is one number the model learned in training. A model of about 1 to 3 billion parameters can be tried on a phone, a laptop, or a board such as a Raspberry Pi. Quantization stores each parameter in fewer bits, often 4 or 8 instead of 16. On one device, record whether the program stayed up, the wait for the first word, and tokens per second. A token is a piece of a word. If the operating system stops the process, the model did not fit in memory.
ONNX (Open Neural Network Exchange) is a file format for a trained model. ONNX Runtime loads that file on a PC or in the browser. The browser build can use WebAssembly on the processor or the WebGPU execution provider on the graphics chip. WebGPU is the W3C standard for GPU work in a web page. The specification marks the API as a secure context, so use HTTPS or localhost. The ONNX Runtime tutorial says Chrome and Edge expose WebGPU on Windows, macOS, Android, and ChromeOS, while support in Firefox and Safari is more limited. Load one model on both providers and write down load time, a second run, and memory.
A legal or finance extractor uses the same log. Ask for parties, dates, amounts, and clause titles, then match each field to a line in the source. Adding a fee that is not printed is a miss. Polaris has no scored public rows for these tests yet. A published row needs the date, machine name, model file, bit width, and the inputs.
Turn one run into a results row
Build a test row from the notes I paste. Do not invent timings, memory figures, or scores that are not in the paste. Return one markdown table with: date, device, RAM, model file, parameter count, bit width, runtime (wasm, webgpu, or other), load result (ran or stopped), time to first token if given, tokens per second if given, and any number in the output that differs from the prompt. Then list the three inputs I should rerun unchanged on a larger model so the rows stay comparable: one receipt or invoice field task, one short speech clip, and one code fix with a test I can execute. Notes: [paste device, model, prompt, and output]
What to write down
Use the same columns for every niche test. Polaris has not published scored results yet. Use these columns for your own test.
| Job | Keep | A miss |
|---|---|---|
| Receipt or invoice | Photo plus true seller, date, total, tax, currency | Wrong field, invented line, or a moved decimal |
| Translation | Source text and a human version | Name, number, date, or unit changed |
| Read-aloud | Sentence and the voice name | Wrong language, skipped words, or silence |
| Speech to text | Audio and a human transcript | Words that differ from the transcript |
| Code assistant | Project, issue, and a test that should pass | Patch that fails that test |
| Small or browser model | Model file, bit width, device, memory | Crash, or a clear drop on the same prompts |
The market and the job board open later. You can run these checks on your own device now.
Before you upload a document or a voice
Before a photo or a clip leaves your machine:
- Cover personal names, card numbers, phone numbers, and any national ID number.
- Keep the original file. A cropped screenshot can hide the line the model misread.
- Write the tool name and the date on the answer key the same day.
- Treat a legal or tax summary as a draft. Compare every amount with the source before you rely on it.
- For speech, say who is allowed to hear the recording. A voice clip can identify a person.
A shared community DGX Station would rerun the heavier speech, receipt-extraction, and code-patch tests on one recorded machine, so a small on-device model and a large model can be compared on the same inputs, date, and machine name. About the shared station
Sources
- Tesseract User ManualOfficial manual for the open-source OCR engine that reads printed text from images. The engine is Apache-2.0. Swedish language data is the swe.traineddata file at https://github.com/tesseract-ocr/tessdata_fast/blob/main/swe.traineddata.
- Translation with built-in AI (Chrome)Chrome's guide to the on-device Translator API, including the language list (Swedish is sv) and the statement that the API runs in desktop Chrome. Docs text is CC BY 4.0; this page does not copy it.
- Web Speech API (MDN)How a web page can speak text (SpeechSynthesis) and, where the browser allows, recognize speech. MDN text is CC BY-SA; this page does not copy it.
- Open ASR LeaderboardPublic table of speech-recognition systems. Use it to see how word error rate and speed are reported, then run your own Swedish clip the same way.
- Open ASR Leaderboard paperThe method paper: word error rate and inverse real-time factor on English and multilingual speech sets. Read the paper for the metric definitions; the live leaderboard is the current table.
- ONNX Runtime WebGPU tutorialHow to run an ONNX model in the browser on the graphics chip, and which browsers the tutorial lists.
- WebGPU specificationW3C Candidate Recommendation for GPU computation in the browser. The API is exposed in a secure context. W3C document license applies to the spec; this page does not copy it.
- SWE-bench leaderboardPublic scores for models that must patch real GitHub issues. The site lists 2,294 full-set instances and 500 in the human-checked Verified subset. Code is at https://github.com/SWE-bench/SWE-bench, which lists the MIT license.