anishfn/shapeshift

An input that becomes what you mean: one text box that morphs into the right UI as you type. Powered by TypeSafe Jev, works offline.

What it solves

Shapeshift provides a single, adaptive input field that automatically transforms into a specialized UI component (a "card") based on the user's intent. Instead of navigating to different tools for timers, checklists, or calculators, the user simply types their request, and the interface morphs into the appropriate tool.

How it works

The system uses a hybrid approach to determine the user's intent:

  1. Intent Classification: It uses the Jev model from TypeSafe AI to classify the input text into one of many categories (e.g., event, checklist, timer) and identify specific signals (like urgency or call type).
  2. Deterministic Parsing: Once the intent is identified, deterministic code handles the actual data extraction (dates, amounts, units) to ensure accuracy.
  3. State Machine: To prevent the UI from flickering as the user types, a state machine manages the transitions between the raw input, ghost previews, and the final committed card.
  4. Offline Mode: A built-in keyword classifier allows the system to function fully offline by default.

Who it’s for

Developers looking to implement adaptive, intent-driven user interfaces or those wanting a versatile, single-input productivity tool.

Highlights

  • Multi-card support: Handles everything from event cards and shopping checklists to bill splitters, color pickers, and time zone converters.
  • Hybrid AI/Code approach: Uses AI for decision-making (intent) and deterministic code for computation.
  • Calam-UI state machine: Implements hysteresis and confidence thresholds to ensure smooth, stable UI transitions.
  • Offline-first: Works without an account or API key via a local keyword classifier.

Related

  • Project
    bilibili/Index-1.9B

    Index‑1.9B is Bilibili’s open‑source 1.9 B‑parameter LLM family, offering base, chat, pure‑instruction‑free, role‑play (RAG) and a 32 k‑token long‑context variant. Trained on 2.8 T tokens of Chinese/English data, it scores competitively on standard benchmarks and excels at processing very long texts. The repo provides ready‑to‑load checkpoints, demo scripts (Gradio, CLI, OpenAI‑compatible API), quantisation support, and a fine‑tuning guide, all under a permissive license with a safety disclaimer.

  • Project
    WorksApplications/Sudachi

    Sudachi is a Java‑based Japanese morphological analyzer that tokenises text, tags parts of speech, and normalises variants. It offers three segmentation modes, large UniDic‑based dictionaries (small, core, full), a plugin architecture, and support for custom user dictionaries. Use it via a command‑line jar, as a Maven library, or through ports in Python/Rust and an Elasticsearch plugin. Licensed under Apache 2.0.

  • Project
    anush008/fastembed-rs

    FastEmbed-rs is a Rust library that generates dense text, sparse text, and image embeddings, and reranks documents, entirely locally using ONNX runtime. It lets developers add semantic search and RAG capabilities to Rust applications without Python dependencies or external APIs.

  • Project
    officecli/officedex

    An AI-native desktop platform that generates native Word, PowerPoint, and Excel files directly from natural language prompts, eliminating manual formatting.