On-device AI tools

Most of these tools use Transformers.js, the JavaScript port of the Hugging Face pipeline API, with ONNX Runtime underneath. On first use the model weights are fetched from the Hugging Face CDN and stored in your browser's cache. From then on the tool works from that cache — the model is on your machine, and re-running it is a local operation.

Execution goes to WebGPU where it is available and the model is suited to it, and falls back to WebAssembly on the CPU otherwise. The difference is large: WebGPU can be several times faster on the image models in particular. Some of the text models here are pinned to WebAssembly deliberately, because they are small enough that the GPU setup cost is not worth paying and because CPU execution is the more predictable path across browsers.

Two tools sit slightly outside that pattern. Image to Text is OCR via Tesseract compiled to WebAssembly, with the worker and core WASM served from this site directly; only the language data file is fetched from the Tesseract project's CDN on first use. Text to Speech uses Piper/VITS voice models — a real neural vocoder producing a downloadable WAV file, not the robotic built-in speech synthesiser — with each voice weighing roughly 20 to 60MB.

One consequence of quantisation deserves a mention because it is visible in the tools. The Translator runs a multilingual model of a few hundred million parameters at 8-bit quantisation, and it is quantised specifically because the full-precision version exhausts browser memory. Quantisation is what makes browser inference possible at all; it also costs a little quality. That trade is present, to different degrees, in every model here.

Audio/Video Transcriber runs Whisper locally. You give it an audio or video file, it decodes the audio in the browser, chunks it with overlapping strides so words are not lost at boundaries, and produces a transcript with timestamps — exportable as plain text, SRT or WebVTT. This is the tool on the site that most clearly justifies the whole category: recorded interviews, therapy sessions, legal depositions and internal meetings are exactly the material people are most reluctant to upload to a transcription service, and it is a base-sized Whisper model, not a toy.

Translator does multilingual translation across sixteen languages with no round trip to a translation API. Text Summarizer runs a distilled summarisation model with short, medium and long presets. Sentiment Analyzer gives a positive or negative classification with a confidence score. Zero-Shot Text Classifier is the more interesting one: you supply your own category labels at runtime, with no training and no fine-tuning, and it scores the text against them — which makes it genuinely useful for triaging a pile of feedback into buckets you invented five seconds ago.

PII Detector & Redactor pairs a named-entity model with pattern matching for emails, card numbers, IP addresses and phone numbers. It is the natural bridge to the AI developer tools: running it locally over text before that text goes to a hosted API is a sensible habit.

One clarification, since it is in the same part of the site: Text Similarity is not a model at all. It is TF-IDF and cosine similarity implemented in plain TypeScript. It is fast, it needs no download, and it is honest about being classical information retrieval rather than an embedding model.

Background Removal is the most immediately useful of the image models — a segmentation model that separates subject from background and exports a transparent PNG, in batches, with the model and its runtime fetched from a version-pinned CDN on first use. Object Cutout is the interactive counterpart, based on a compact Segment Anything variant: click a point on the thing you want, click again to add or exclude regions, and get a cut-out with a transparent background. Click-to-select is a much better interface than a lasso tool for anything with a complicated edge.

Image Upscaler runs a super-resolution transformer at a fixed 2x. It is real reconstruction rather than interpolation — it does better than bicubic on text and edges — and it is not the marketing version of upscaling: there is no invention of detail that was not implied by the input, and the input is capped in size because the memory cost scales with it.

Image Captioner produces a one-line description of a photograph, which is useful for draft alt text over a batch of images, with the emphasis on draft. Depth Map Generator estimates per-pixel depth from a single photo and renders it as a greyscale or false-colour image — input to compositing, parallax effects, or 3D work, from a plain photograph and no depth sensor.

Use on-device inference when the data is sensitive, when you have volume and no budget, when you need it to work without an account, or when the task is genuinely small — a sentiment classification does not need a frontier model.

Use a hosted API when you need the best available quality, when the input is long and the latency matters, when you are running this in production for other people, or when the task requires reasoning that a 100-million-parameter model simply cannot do. The AI developer tools collection is the set of utilities for that path.

The two are complementary more often than people assume. Redact locally, then send. Transcribe locally, then summarise with a hosted model. Classify locally to route, then escalate the hard cases.