Axiophase
Frontier AI engineering

We build what
comes next.

Axiophase is an AI engineering lab. We build real-time video & ML pipelines, AI agents and the cloud behind them — and release what we make as free software.

See our work → Read the blog
// Projects

Things we've built.

Free software, real constraints. Every project is a chance to push what runs on a phone, a camera or a cloud.

Tong Learn icon Free · Android

Tong Learn

An offline English ↔ Mandarin and Cantonese translator for travellers. Speech recognition, translation, a 1.5B-parameter language model and camera text reading — all running on the phone, no internet needed.

View project → How it's built

More free apps

In progress

Travel Talk

Tong Learn's engine taken multilingual: Japanese, Korean and the major European languages, each a downloadable offline pack.

  • Rust core · Kotlin · on-device LLM
In progress

Travel Companion

Imports your itinerary, follows it offline during the trip and proposes safe changes when plans change — never touching fixed bookings without asking.

  • Rust · Android · AI planning

Video & ML tooling

Video & ML

Edge video analytics

Real-time detection and tracking on NVIDIA Jetson: RTSP decode, TensorRT inference, a Rust temporal-tracking core with Kalman prediction and Hungarian matching, and event analytics replayable against recorded footage.

  • Rust · TensorRT · DeepStream · ZeroMQ
Video & ML

Pipeline Studio

Visual debugger for inference pipelines: build the graph, run it with cached frames, compare two detectors side by side and mine the frames where they disagree.

  • FastAPI · React Flow · Parquet · SQLite
Data

YOLO annotation workspace

Labelling for huge local image datasets. Labels stay as plain files on disk; thumbnail-first browsing, box propagation across frames and similar-scene search.

  • Next.js · Rust/Axum · FastAPI · CLIP
Data

DataForge

Dataset curation on a UMAP map of CLIP embeddings: cluster, lasso-select, keep, remove or relabel thousands of samples, with sub-millisecond neighbour search.

  • React · FastAPI · DuckDB · HNSW
Data

Classifier review

High-speed review of classifier crops: big thumbnails, drag-to-bucket, bulk moves with undo, and an accept gate that writes a clean dataset snapshot. Spawned automatically by the mining pipeline.

  • Next.js · NestJS · SQLite
Video & ML

ROI & zone tooling

Draw, extract and scale exclusion-zone polygons for camera views — including pulling zones from customer diagrams and checking them against live frames.

  • Python · OpenCV · LabelMe

Agents & operations

AI agents

Academic Strategist

A team of LLM agents that plans years ahead for a student: a lead strategist consulting specialist agents, long-term memory and follow-ups on its own advice.

  • Multi-agent · memory · scheduling
DevOps · Agents

Edge AI fleet operations

Runbooks an AI coding agent can execute: deploy sites to Jetson hubs, export models to TensorRT, verify cameras and events, recover offline devices and keep disks and logs in check.

  • Jetson · TensorRT · Docker · agent skills
// Expertise

Deep in the stack,
end to end.

From model to silicon to cloud — we design, build and run systems that have to work in the real world.

Video & ML pipelines

Real-time inference on video and audio streams, optimised for GPUs at the edge and on phones. Training, export, quantisation and deployment.

  • DeepStream, GStreamer, TensorRT, ONNX Runtime
  • YOLO detection, tracking, OCR, speech
  • On-device LLMs with llama.cpp

AI agents

LLM agents that reason, use tools and act over long time spans — with memory, evaluation and guardrails built in.

  • Tool-calling & multi-agent orchestration
  • Retrieval and long-term memory
  • Evaluation, monitoring, cost control

Cloud

Scalable, secure, cost-aware infrastructure for ML workloads: GPU inference, storage and event ingestion.

  • AWS, GCP
  • Kubernetes, serverless, GPU inference
  • Infrastructure as code

DevOps

Reproducible builds, continuous delivery and observability — for cloud services, edge device fleets and mobile apps alike.

  • Docker, CI/CD, GitOps
  • Logging, metrics, alerting
  • Remote fleet deploy & recovery
RustPythonKotlinPyTorchTensorRTDeepStreamONNX Runtimellama.cppsherpa-onnxPaddleOCRDockerKubernetesTerraformAWSJetson
// Blog

Notes from the build.

What we measured, what worked and what didn't.

An app that can't phone home

Why Tong Learn doesn't ask for internet access at all — and what that costs.

→

Reading a menu photo in 1.8 seconds on a phone

Thread counts, OCR and why recognition, not detection, is the bottleneck.

→
All posts →
// About

Built by engineers who ship.

Axiophase is a small AI engineering lab. We've run computer vision on edge fleets at live industrial sites, put AI agents into production and operated the cloud and DevOps behind them. Today we build free software to explore what's possible — and we'll launch products when the right idea earns it.

Edge→Cloudone stack, end to end
On-deviceAI that works offline
Real-timeaudio & video AI
Freesoftware we release
// Contact

Say hello.

Questions about our work, feedback on Tong Learn, collaboration, press or careers — we'd love to hear from you.

hello@axiophase.com