Staff Engineer, AI Harness

Atmospheric AI Limited; Greater London; Hybrid; GBP 110,000 - 150,000; Full time

Job summary

Atmospheric AI Limited is building a ground-truth economics and procurement platform that models material costs and carbon exposure. You will design a scalable tool layer, orchestrate modular AI agents, and ensure context is governed with traceable sources.

You’ll implement evals, guardrails and performance benchmarks to outperform general AI in procurement tasks. The role blends engineering depth with product impact, requiring strong Python and frontend capability to land results in UI

Qualifications

• Built high-scale platforms for enterprise clients.
• Created dependable AI systems with verifiable evals and guardrails.
• Led end-to-end ownership of technical domains and raised team capability.
• Iterated rapidly with practical delivery in days.
• Daily use of AI to drive tooling and automation.

Responsibilities

• Expose physics model as composable tools that others build against.
• Design orchestration of models and how work is decomposed and run.
• Assemble and ground context with sources within budget.
• Ship and evaluate modules with clear performance benchmarks.
• Define runtime guardrails for cost, latency and failure modes.
• Lead future work into specialised models as trust grows.

Skills

Python

TypeScript

React

Tools

Dagster

Postgres

FastAPI

AWS

LangGraph

Pydantic AI

Next.js

Job description

Everything physical is made from raw materials: polymers, chemicals, metals and minerals, a $16T economy that software has barely modelled. Atmospheric AI computes what those materials should cost, what carbon they carry and where their supply is exposed, from the physics of how they're made, so manufacturers can decide and negotiate from ground truth.

The problem

Every company has the same frontier models. What makes one fit for a decision with money on it is the harness: the tools, context and checks that ground a model in a domain. A procurement lead asks whether a supplier's surcharge is justified, or what happens to her polymer cost if crude moves. Cortex has to answer with a number that is computed, traceable and defensible in a negotiation. You lead it: design, plan and execution.

What you'll own

• The tool layer. Expose our physics model as composable tools agents can orchestrate, and set the standards other engineers and agents build against.
• Agents and orchestration. Models are swappable; the logic lives in the harness. You own how work is decomposed, routed and run.
• Context, memory and grounding. Context assembled when it's needed, kept within budget, and each figure tied to its source.
• Evals and the benchmark. Every module ships through evals. You prove, module by module, where and by how much our system beats general-purpose AI on procurement work.
• Runtime and guardrails. Cost, latency and failure modes designed for up front, so wrong answers can't ship silently.
• What comes next. As trust is earned, the harness moves from answering to acting, and where it pays, you lead us into specialised models of our own.

What we're looking for

• You've built high-scale platforms that enterprise clients depend on. Walk us through one you built and why it worked. That's where every conversation with us starts.
• You've built the machinery that makes AI dependable. A production AI or agent system where the logic lived in the system, context was engineered, correctness was measured with evals and failure modes were designed for.
• You multiply. You've owned a technical domain end to end and raised a team's ceiling through the systems and standards you built.
• You iterate in days. Ship, measure, tighten. Expert Python, and enough TypeScript and React to land a result in the UI.
• AI is your daily leverage. You direct agents in your own workflow and would rather build the tool than repeat the task.

Nice signals: durable workflow orchestration; serious tool-use or function-calling work; retrieval beyond tutorials; applied AI where wrong answers have consequences; supply-chain, commodities or manufacturing exposure.

Stack: Python, FastAPI, Postgres, Dagster, AWS; LangGraph, MCP, Pydantic AI; Next.js and TypeScript. A signal, never a requirement.

The take-home: a short task shaped like the real work, and later a working session on a problem like ours.