Features

A short tour of what Pipelogic ships, with links to the in-depth pages for each area.

Typed dataflow

Every connection between components carries a typed stream. The type system is enforced at backend-wire time — connecting an Image output to a [BoundingBox] input fails immediately, before anything runs.

The type catalog covers atomics (Int32, Bool, String, Bytes, etc.), collections ([T], (T, U), records, unions), and named domain types (Image, BoundingBox, AudioFrame, Tensor, Mask, Landmark, Polygon, VideoFrame, ...).

The component model

A component is a small program with a typed I/O contract. It can be written in Python (pipelogic package, fast iteration, ML-friendly) or C++ (pipeml library, low latency, full Triton/tracker access).

Pipelogic ships 165+ components covering computer vision, audio, TTS, OCR, LLM access, messaging, file I/O, and visualization. You can use them as-is, fork, or write your own.

Stream transformations

21 built-in transformations let you reshape data inside a backend without writing a custom component:

• Cardinality: flatten, lift_unroll, lift_reroll, length, repeat, infinite-repeat
• Joins: join, shuffle, unite_streams, select_stream
• Type conversion: convert_value, constant
• Pack / unpack: tuple, record, named, union (pack_* / unpack_*)
• Predicates: filter, cond, delay_by_one

Model deployment

Components that need ML inference connect to a serving runtime. Five ship by default; you can add your own:

Runtime	Best for
Triton	ONNX, TensorRT, PyTorch, TensorFlow
TorchServe	PyTorch models with custom handlers
Ollama	Local LLM inference
vLLM	High-throughput LLM serving
SGLang	Vision-language models, structured generation

The platform manages the runtime lifecycle — your component just declares which runtime it needs in component.yml and the platform spins it up.

Visual editor + CLI

The App lets you design backends visually with live preview. The ppl CLI lets you do everything from a terminal — build, release, deploy, debug, monitor.

Both manipulate the same underlying backend objects. Anything you build in one is editable in the other.

Managed cloud, on-prem, or air-gapped

Pick the runtime that matches your constraints — managed cloud for elastic GPUs, on-prem for your own hardware, air-gapped for sites with no internet egress. The same backend deploys unchanged onto any of them.

Bring your own everything

The bundled model-serving runtimes (Triton, TorchServe, Ollama, vLLM, SGLang) cover most ML inference. Everything else — third-party APIs, message brokers, databases, browsers, hardware devices — plugs in by writing a typed component. Same SDK, same typed streams, same ppl release. There is no separate plugin or extension API.

Maturity at a glance

Feature	Maturity
Web backend editor	🟢 Stable
ppl CLI	🟢 Stable
Python SDK (pipelogic)	🟢 Stable
C++ SDK (pipeml)	🟢 Stable
Live streaming, multimodal	🟢 Stable
Stream transformations	🟢 Stable
Triton, TorchServe, Ollama, vLLM, SGLang serving	🟢 Stable
Backend-building agents	🟡 Preview