About Us
PrismML builds high-performance foundation models designed to run efficiently across a wide range of environments—from edge devices to large-scale deployments. Our work spans models from ~1B to 100B+ parameters across LLMs, diffusion models, and other modalities, with a strong focus on scalable training, efficient inference, and real-world deployment. Today, that means a 27B-class model that runs on a phone in about 4 GB, down from roughly 54 GB at full precision.
Our Bonsai family of 1-bit and ternary models is designed to dramatically improve the efficiency of modern AI systems, enabling advanced intelligence to run with significantly lower memory usage, latency, and energy consumption across cloud and edge environments.
Role Overview
We are seeking a Senior-level (or higher) AI/ML Engineer to design and build the agent harness and runtime that turn Bonsai models into reliable, useful agents. This role focuses on the systems surrounding the model: the execution loop, tool interfaces, context and memory management, task orchestration, and evaluation infrastructure that enable agents to complete complex workflows.
You will own the harness from architecture through production, working closely with model, infrastructure, and product teams. A central challenge is making efficient models effective at real tasks: determining what the model needs from its environment, measuring where execution breaks down, and improving the system across cloud, local, and resource-constrained deployments.
Responsibilities
You will design, build, and optimize the systems that enable Bonsai-powered agents to perform real work. Core responsibilities include:
- Developing the core agent execution loop, including model invocation, tool selection and execution, structured outputs, stopping conditions, and error recovery
- Building context and memory systems for long-running tasks, including retrieval, context compaction, persistent session state, and reusable skills
- Designing reliable task orchestration with checkpointing, cancellation, retries, and resumable execution, including coordination across multiple agents where appropriate
- Building extensible tool interfaces and integrations with APIs, browsers, terminals, filesystems, and external applications
- Implementing execution safeguards, including sandboxing, scoped permissions, credential handling, defenses against prompt injection from untrusted content, and human approval for sensitive actions
- Developing task-based evaluations, regression suites, execution traces, and debugging tools to measure completion quality and diagnose failures
- Optimizing task success, latency, token usage, memory consumption, and cost across model configurations and deployment environments
- Collaborating with research and post-training teams to turn agent failures and execution traces into actionable model improvements, and to reuse harness tasks as reinforcement-learning environments
Basic Qualifications
You bring experience building reliable agent systems and production LLM applications:
- 5–8+ years of experience in software engineering, ML systems, distributed systems, or developer infrastructure
- Strong programming skills in Python and experience building maintainable services, APIs, and developer tooling
- Hands-on experience building agents or LLM applications involving tool use, multi-step execution, and state management
- Solid understanding of concurrency, asynchronous execution, persistence, and failure recovery
- Practical understanding of LLM behavior, including context limitations, tool calling, structured outputs, and the interaction between prompts, models, and execution environments
- Ability to take an ambiguous technical problem from initial design through implementation, deployment, and iteration
- Strong written and verbal communication skills
- Proven ability to mentor and collaborate effectively with other engineers
Preferred Qualifications
You have additional experience aligned with agent systems and efficient model deployment:
- Experience building or contributing to agent harnesses, runtimes, or frameworks such as Hermes Agent, Codex CLI, OpenHands, or comparable open-source systems
- Experience with durable workflow engines, sandboxed code execution, container runtimes, or secure tool execution
- Familiarity with inference runtimes such as llama.cpp, MLX, or vLLM, and integrating locally hosted or open-weight models into agent systems
- Experience making smaller or quantized models reliable in agent loops, including constrained decoding, grammar-based structured outputs, and tool schemas and context budgets adapted to model capability
- Experience building agents that run fully on-device or offline, without a cloud fallback
- Experience designing agent benchmarks, task environments, automated graders, or evaluation methods that account for nondeterministic behavior
- Familiarity with model routing, prompt and prefix caching, and tradeoffs between model capability, execution speed, and cost
- Contributions to open-source AI infrastructure, developer tools, or technically substantial independent projects
Ideal Candidate Profile
You have built systems that keep working when models make mistakes, tools fail, and tasks take longer than expected. You can move between inspecting an agent trace, improving context selection, and debugging a runtime failure. You use measured task outcomes to guide architectural decisions and know when to introduce an abstraction or simplify one.
You are excited by the opportunity to make highly efficient models useful across a much wider range of devices and workflows. You bring strong systems judgment, clear technical opinions grounded in experience, and the ownership to turn an early harness into infrastructure that developers and users can depend on.