AI & ML interests

None defined yet.

Recent Activity

asermnasr  updated a model about 9 hours ago
figranium/folate
asermnasr  updated a Space 1 day ago
figranium/folate-demo
asermnasr  published a Space 1 day ago
figranium/folate-demo
View all activity

Organization Card
Figranium

Open-weight machine learning models engineered for low-latency browser automation and deterministic DOM grounding.

Figranium   ·   GitHub   ·   Community

Figranium develops open-weight machine learning architectures engineered for deterministic browser automation, structural DOM grounding, and low-latency web-agent inference.

This Hugging Face organization serves as the official repository for pre-trained model weights, quantized inference artifacts, and evaluation tooling released by Figranium. Rather than relying on high-latency Vision-Language Model (VLM) round-trips or cloud-bound API dependencies, architectures published under this namespace are designed for local, deterministic execution on standard CPU runtimes.

Overview

Modern web automation pipelines frequently suffer from high inference latency, token overhead, and non-deterministic selector resolution when routing raw screenshots or unpruned HTML through general-purpose foundation models. Figranium addresses these constraints by training compact, task-specific architectures that operate directly on pruned structural DOM representations and semantic attributes.

Key engineering priorities include:

  • CPU-Native Execution: Models are quantized to INT8 and optimized for minimal memory footprint and sub-second latency on commodity CPU containers without requiring GPU hardware acceleration.
  • DOM-Centric Grounding: Inference pipelines evaluate structured candidate nodes, accessibility attributes, and hierarchical context rather than high-resolution visual frames.
  • Self-Hosted Reproducibility: Weights and tokenization configurations are distributed openly to integrate directly into self-hosted automation runtimes and deterministic execution engines.

Models

  • Folate is an INT8-quantized CPU DOM ranker designed for high-precision UI element localization. Given a natural-language action intent and a pruned set of candidate DOM elements, Folate computes relevance scores across interactive nodes to identify target elements in sub-second timeframes. Built to replace heavy multimodal grounding loops, Folate executes locally within headless browser environments and containerized automation workers.

Integration

Models published under this organization integrate natively into Figranium, an open-source browser automation platform and visual workflow engine powered by Playwright, providing local inference for automated selector resolution and action execution.

Usage and Licensing

All model weights and configuration files are hosted under their respective repository licenses. Refer to individual model cards for specific architecture specifications, input serialization schemas, quantization benchmarks, and inference examples.

datasets 0

None public yet