pbj

Build fast agents,
faster than ever.

Copilotserve

Fine-tune with control

Drag, drop, train. Go from idea to optimised model without writing a single line of code.

Zero-code Fine-tuning

Build complete training pipelines visually. Connect agents, rewards, and datasets, code-free.

Trainable
Mark agent for training
Trainable
×
Agent

Research-grade Methods

Implement state-of-the-art research in minutes.

Agent
Agent
Reward
Reward

Any Model, Any Size

Swap between open-weight models in one click - Llama, Qwen, Gemma, Mistral, or bring your own.

Qwen 2.5
7B
Llama 3.1
8B
Gemma 2
9B

One-click Cloud Training

Hit train and we handle the GPUs, checkpoints, and monitoring - no code, no boilerplate.

Why PBJ

pbj learns with you

As your platform scales, so does pbj's knowledge about how to make you more efficient, enabling continuously optimized compilation across the stack.

Create
Observe
Mine
Optimize
Flywheel
Build flows by hand or with copilot

You create agent flows by hand or with our AI copilot. Our diagrammatic representation lets pbj automatically optimize your flows, offering suggestions with annotated production performance to discover the most efficient flows for your use case.

The result is your own self-growing moat of optimizations for better performance and lower costs.

No other platform combines copilot-led visual orchestration with multi-agent finetuning to learn with you.

Proprietary Optimizations

Research-backed techniques that compound with scale.

Simulated Finetuning

~$0.01~$1,000s

Cheap runtime finetuning, model-size agnostic.

Efficient Swarm Protocols

5–15%cheaper per forward pass

Training-free optimization for swarms of small-medium models.

Long-Context Inference

5xspeed + 80% per-token savings

Maintain perplexity while slashing costs on long contexts.

83%
Total cost reduction
20% per token × 85% per forward pass = 17% original cost

Future full-stack optimizations

Intelligent KV cache management for shared prefixesPrefetch schedulingOptimal prefix retention estimationPer-swarm execution trace optimizationHierarchical texture gradients

Built by Engineers, for Everyone.

Our team brings together expertise spanning ML research, kernel engineering, and compilers. We don't just integrate existing tools; we build the underlying infrastructure that makes fine-tuning and serving genuinely fast.

Compiler Design

We compile visual graphs into clean, debuggable Python so your pipeline remains transparent and hackable.

Kernel Engineering

Runtime kernels are tuned for memory and throughput, so training jobs scale efficiently across mixed hardware.

ML Research

We build and validate new optimisation strategies to improve convergence speed and model quality.

Production Systems

From autoscaling to rollout safety, we design deployment paths that are reliable in real-world traffic.

Team

Scientists and engineers with experience across the stack - from leading organisations like Oxford, Arm, Quantinuum, Morgan Stanley, and LinkedIn.

Blake Wilson profile

Blake Wilson

Research Scientist & CEO

PhD in Electrical and Computer Engineering, Purdue University


  • SoC Engineer, Arm and Purdue SoCET
  • AI for Physics & Optimization, ML Acceleration
Nikhil Khatri profile

Nikhil Khatri

Research Scientist

DPhil Candidate, Machine Learning Research Group, University of Oxford


  • ML for Predictive Maintenance, Data Centers, LinkedIn
  • ML Theory, Systems Engineering
Jules Desai profile

Jules Desai

Research Scientist

MPhysPhil Physics and Philosophy, University of Oxford


  • Oxford Cosmos Fellow & 2x Gibbs Prize
  • Multi-Agent Systems, Human-Centric AI, Electronic Musician
Thomas Weatherbee profile

Thomas Weatherbee

Research Scientist

MMathPhys Mathematics and Theoretical Physics, University of Oxford


  • Quant, Squarepoint & Morgan Stanley
  • Multi-Task Learning, Agent Swarms, ML for Finance
Lukas Heidemann profile

Lukas Heidemann

Research Scientist

DPhil Computer Science, University of Oxford, MSc Mathematics, University of Bonn


  • Compiler Engineer, Quantinuum
  • Performance Engineering, Programming Language Theory, Homotopy Theory
Vincent Wang-Maścianica profile

Vincent Wang-Maścianica

Research Scientist

DPhil Computer Science, University of Oxford


  • Senior Researcher, Oxford & Quantinuum
  • Category Theory, ML Theory
Marcus Gemzøe-Winding profile

Marcus Gemzøe-Winding

Research Scientist

MSc Computer Science at DTU


  • CTO, Spher
  • Full-Stack Systems, ML Infrastructure
Sri Tirukkovalluri profile

Sri Tirukkovalluri

Research Scientist

MSc in Mathematical & Theoretical Physics, University of Sheffield


  • Harvard GSAS Assistant
  • RL for Quantum Chemistry, AI for Physics, Nanophotonics

Affiliations & Partners

University of OxfordQuantinuumQuEraPurdue UniversityDTUSquarepointUniversity of CambridgeUC BerkeleyOak Ridge National LabUniversity of OxfordQuantinuumQuEraPurdue UniversityDTUSquarepointUniversity of CambridgeUC BerkeleyOak Ridge National Lab

Platform Comparison

pbjOpenAI Agent BuilderLangflowLangGraphTinkerModalTRL
VisualYesYesYesNoNoNoNo
EaseEasyEasyEasyModerateExpertModerateExpert
Multi-agentYesYesYesYesN/AN/AN/A
Fine-tuneYesNoNoNoYesYesYes
HostingYesYesPartialNoYesYesNo
ModelsAnyGPT onlyAnyAnyAnyAnyAny

Start from an Example

Open a pre-built pipeline in the editor to see what's possible, then tweak it to fit your use case.