Staff Engineer, Core ML
Confirmed on the employer's own hiring board on Nov 10, 2026. First seen by Alion on Nov 10, 2026.
About Circuit
Circuit is the AI execution layer for the physical economy.
Industrial and manufacturing enterprises run on expertise: the hard-won knowledge of how complex products are engineered, configured, sold, and supported. That expertise is scarce, slow to transfer, and walking out the door as a generation of experts retires. It lives in PDFs, spreadsheets, and the heads of a few people, and the hardest technical decisions wait on whoever happens to know.
Circuit puts expert judgment in everyone's hands. We capture how the best people make critical technical decisions and turn it into agents that do the judgment-heavy work, so anyone, regardless of skill level, can make the right call, from quote to install and support, with the accuracy these environments demand.
Circuit is built by operators. Our leadership team has built and scaled real industrial companies, and we are already live in production with category-leading manufacturers. We are not bolting AI onto an old playbook. We are building the layer that runs the last mile of how the physical economy executes.
The Role
As a Staff Engineer, Core ML at Circuit, you’ll join Product Engineering and own the ML capabilities at the center of our product: understanding complex technical documents, and keeping the long-running agents that work with them reliable.
Our customers’ knowledge lives in blueprints, CAD files, spec sheets, and scanned PDFs that mix text, vector geometry, and raster content. You’ll build the models and algorithms that turn those documents into structured information, and the tools that let our agents use it. You’ll also improve the agent loop itself: how agents plan, choose tools, recover from failure, and behave consistently from one run to the next.
Most of your time goes into the algorithms and models behind the product rather than individual customer projects. Product engineers and FDEs work closest to customers and bring you the failure patterns they see. You analyze how our models and agents perform, decide where better algorithms, data, or training will pay off, and set the evaluation bar that tells us whether a change worked. You’ll join customer conversations when a problem needs your depth.
This is an applied role, not a research one. You’ll adapt and combine proven approaches, measure them against real customer data, and ship them. Your teammates range from full-stack engineers to applied ML researchers, and success is measured by what reaches production, not papers published.
What Is in Production Today
Our core product includes a long-running agent loop written in Go. Frontier model APIs drive the loop, while specialized tools handle retrieval, extraction, and agent execution. These tools let us solve industrial problems that general-purpose agent harnesses struggle with.
We use a combination of open-source and internally fine-tuned models, particularly vision-language models (VLMs), hosted on KServe. Our customer data includes technical documents, high-resolution blueprints, raster images, vector geometry, CAD, and structured data.
Key Responsibilities
Develop algorithms that extract meaning from complex technical documents containing text, vector geometry, CAD, and raster content, spanning vision models, geometric algorithms, and language understanding.
Build those document understanding capabilities into the tools that power Circuit's long-running agentic systems.
Improve the reliability, determinism, and effectiveness of agentic systems, including agent planning, tool use, and failure recovery.
Build and extend training and evaluation pipelines for our core models and agents, so every change is measured against real customer data before it ships.
Analyze how our algorithms perform in production, find where and why they fail, and turn that into the next improvement.
Find gaps and bias in our datasets, and close them through curation, labeling, or synthetic data.
Track new research and techniques, and make the promising ones work in production.
Mentor and elevate other engineers on the team.
Examples of Work You Could Own
Fix a failure class. Trace why extraction breaks on a class of drawings, such as dense electrical schematics, fix it with better data, a better model, or a better algorithm, and show the lift in the evals.
Fine-tune a specialized model. Fine-tune a LoRA on an in-house model to extract a specialized artifact from a high-resolution blueprint page.
Bring a new algorithm to production. Take an approach from a recent paper, such as a new layout or table-structure model, prove it beats what we run today on our data, and make it fast and reliable enough to ship.
Improve retrieval. Improve retrieval quality, relevance, or latency across document types and customers.
Make agents more deterministic. Find why the same task produces different results from run to run, and change the tools or the loop so it stops.
Experience
Agent harness design and execution
Document extraction and structured data understanding
High-resolution computer vision
VLM fine-tuning and inference
Retrieval systems, including embeddings and reranking
Agent evaluation and optimization
Hands-on AI/ML engineering experience, with a track record of shipping models in production.
Experience training and fine-tuning models for production, including the datasets and pipelines behind them.
Experience building and operating ML evaluation pipelines, including datasets, metrics, and regression testing for models and agents in production.
Experience profiling data distributions, identifying gaps and bias, and curating high-quality datasets for training and evaluation.
Experience with vision and multimodal models, especially on technical documents (PDFs, CAD) with text, vector geometry, and raster content.
Strong knowledge of data structures, algorithms, and software engineering principles.
Proven expertise with LLMs, deep learning systems, and applied ML.
Experience debugging production ML systems using AI observability tools (e.g. Langfuse, Opik, etc.), tracing failures to root cause.
Proficiency in Python and relevant ML frameworks (PyTorch, vLLM, etc.).
Daily, fluent use of AI coding tools. You write effective prompts, iterate when output is wrong, and reliably tell working code from code that only looks right.
Bonus: Computational geometry, or experience with vector drawing formats such as CAD and vector PDFs.
Bonus: Go, the language our agent loop is written in.
- Bonus: Familiarity with construction, civil engineering, or related industries where interpreting complex technical drawings and specifications is central to the work.
We want someone who has shipped ML-powered software to production, can adapt proven approaches and evaluate alternatives from other production ML systems, and measures whether a change works for customers. Ideally, you have deep expertise and strong technical opinions in at least one of these areas:
You do not need to be an expert in every area. Strong software engineering fundamentals and the ability to lead technical decisions matter more than familiarity with our exact stack. Beyond that, we look for:
Human Skills
Ownership mindset with a focus on delivering working systems in production.
Strong product judgment, with an understanding of how system behavior impacts user experience and trust.
Rigorous reviewer of agent instructions, skills, and documentation, ensuring they’re clear, unambiguous, and reliable for agents to act on.
Curiosity about how systems behave in practice and how customers use them.
Bias toward shipping. You take ambiguous problems and drive them to working solutions without needing the path fully defined for you.
Judgment about what is ready to ship, and the willingness to push back to keep the work grounded in customer outcomes.
Clear communicator. You explain ML tradeoffs to engineers, FDEs, and product stakeholders in plain terms, and share progress early.
Critical evaluation of AI output. You spot hallucinated APIs, incorrect logic, and subtle bugs before they reach production.
Systems thinking, including attention to correctness, determinism, and failure modes in long-running agents.
Rigor about measurement. You measure before you claim an improvement, and you can tell a real gain from noise in an eval.
Raises the bar for the team through design reviews, code reviews, and mentoring.
Low ego, with a focus on team outcomes.
What We Offer
High-caliber team of seasoned manufacturing and AI leaders
Help create a category-defining platform transforming how manufacturers sell, support, and innovate
Competitive compensation, meaningful equity, and employer-paid benefits
Flexible PTO and access to the latest AI tools
A values-driven culture that rewards ownership, curiosity and velocity