general tensor

We're a team of ml engineers (x-google, x-apple) based in san francisco. We've develop small, specialized models that do exceptionally well at specific tasks that large, general-purpose models suck at. Tasks that demand precision, speed, and reliability over broad capabilities.

List of our models:

pii-redactor-general-V2

security

A compact vlm combined with rules to identify and de-identify personally identifiable information in text and images

code-diff-merge-V2

code

Fast and accurate small language model for applying code diffs

signature-detect-V1

content

Vision model to identify and verify handwritten signatures in images

toxic-guard-text-V1

safety

Real-time detection of harmful language and content in text

toxic-guard-image-V1

safety

Real-time detection of harmful language and content in images

embedding-base-V1.5

batch

High-throughput multi-modal encoder for large-scale processing

logo-locator-V0.1

content

Vision model to detect and locate logos in images

jailbreak-guard-V1

security

Protection against prompt injection and jailbreak attempts

voice-detect-V1

audio

Small audio model for turn detection and speaker verification

prompt-injection-shield-V1.1

security

Real-time detection and prevention of prompt injection attacks with a classifier powered by a database of embeddings

nsfw-image-classifier-V1

safety

Image classifier for not-safe-for-work content

financial-sense-V0.1

safety

Text classifier for detection of sensitive financial content (e.g. financial advice) in text and images

ocr-table-extract-V1

content

Compact multimodal model for extracting tabular data embedded in scanned documents and pdfs

text-autocomplete-V1

content

Small language model for fast, accurate context-aware text completions

tts-alignment-V1

audio

Precise text-to-speech timing and phoneme alignment

typo-correction-V1

content

Tiny language model for spelling and grammar correction

noise-suppression-V0.2

audio

Real-time noise reduction and cleanup for noisy audio

embeddings-text-log-V0.1

code

Embeddings for log analysis and search

pipe-sql-V0.2

code

Small language model to convert NL queries to SQL statements

object-tracking-video-V0.2

content

Track custom objects in video

pii-redactor-name-V1

security

Small language model to identify and de-identify names in text

What makes our models different:

  • Task-optimized: each model is built for one thing, does it exceptionally well
  • Fast: 5-100x faster inference than general models
  • Cost-efficient: also radically cheaper
  • Lightweight and flexible deployment: simple API access or runnable on edge hardware -- no torch or other bulky dependencies
  • Safe and performant: we regularly review, audit, and benchmark all models for continuous improvement

contact

We're currently in private beta. Contact us if you're interested in using our models.