Staff Software Engineer (Agentic Analytics)
Posted Updated
Description
- We’re building the foundational infrastructure that will define how the world thinks about and deploys AI, and we want the sharpest, most driven, and most curious people to help us do it
- As the founding lead of Agentic Analytics, a new team in our Analytics and Data Insights organization, you’ll own how Cohere gets answers from its own data: an agentic analytics system that answers business questions over governed data, shows the source, query, and freshness behind every answer, and asks for clarification or declines when the evidence is insufficient
- Agentic analytics is far from a solved problem, and correctness is the hard part. A query that runs can still answer the wrong question. Agents go wrong when a business term maps to more than one table, when data goes stale, and when the right context never reaches the model
- The worst failures are silent and plausible: a join that double-counts, a wrong denominator, missing data read as zero. Getting this right takes governed definitions, context an agent can use, and an evaluation bar that decides what ships
- You’ll build on work already in motion: an internal analytics agent in pilot on our BigQuery warehouse, which is fed by tested ingestion pipelines and modeled in dbt
- In your first year, you’ll take it to production for teams across Cohere, one domain at a time, and make it replicable: each new domain stands up with configuration, a fresh evaluation set, and very little new code
- This is a hands-on staff engineering role reporting to our Senior Director of Data Analytics. You’ll write production code and evaluation sets yourself while managing a small founding team that grows as the system proves itself
- Own the system end to end. Set the architecture for how a business question becomes a governed query and a verifiable answer: interpreting the question, retrieving the right context, applying approved definitions and joins, enforcing the asking user’s permissions, and attaching the evidence (source, query, freshness) to every answer
- Build and run it in production. Design the core services and write much of the first version yourself: agent-facing data tools (MCP or equivalent), context retrieval, agent skills for analytics, and the APIs that serve answers
- Hold them to production standards for reliability, latency, cost, security, and observability. Put it where people already work: Slack, North, IDEs, and other agents
- Hold every release to the evaluation bar. Build and run the evaluation harness, and define what correct means in each domain: ground-truth question sets, including held-out cases and questions the agent should clarify or decline; result-level scoring, correct-or-decline rates, coverage, and error severity; and regression tests on every change to context, skills, or models. Nothing ships below the bar, and the call on when a domain is ready for production is yours
- Engineer the context layer. Turn metric definitions, approved joins, documentation generated from the code that builds each table, a business glossary, and worked examples into context the agent can use. Every correction becomes a reviewed definition and a regression test
- Roll it out across Cohere. Bring domains such as marketing, revenue, finance, and product onto the system one at a time, each with named owners, certified definitions, and its own evaluation set. Measure adoption, accuracy, turnaround, and correction effort against how those teams work today
- Make it replicable. Package the tools, skills, context templates, and evaluations, and document them well enough that other teams at Cohere can run the system without you
- Hire and lead the team. Bring on the team’s first engineers, set the technical bar, review the work, and develop the people doing it
- Partner across Cohere. Work with our data platform team on the warehouse and semantic models, our model evaluation team on evaluation methodology, our security team on access, and each domain’s owners on definitions and rollout
Benefits
- Six weeks’ paid vacation
- Equity / stock options
- RRSP, 401(k), and Pension Scheme contributions
- Coverage for 100% of your insurance premiums across health, dental, vision, and travel
- Additional coverage for accessing mental health providers/services
- Six months of fully paid parental leave, including adoption and surrogacy
- Financial support for egg freezing and IVF in Canada and the UK
- A monthly fitness and wellness allowance
- Globally dispersed company that supports a remote work culture
- A $2,000 annual education benefit for professional development
- A weekly stipend for meals when working remotely and catered lunch when working from one of our global offices
- A monthly arts and culture allowance
- A monthly quality time allowance
- You may have held the title software engineer, AI engineer, or data scientist; what matters is that you’ve built a system like this and run it in production
- Experience shipping LLM agents to production: tool use, MCP or similar, retrieval and context engineering, guardrails, and observability, with firm opinions on where they fail
- Strong Python and SQL, and experience designing and running production services and APIs on a major cloud (Google Cloud is a plus)
- Experience building identity-aware access for agents, where the agent acts with the asking user’s permissions and every query leaves an audit trail
- Depth in evaluating LLM systems: golden sets, result-level comparison, abstention metrics, error taxonomies, and regression testing in CI, plus the judgment to tell a good demo from a system you’d put in front of a CFO
- Hands-on experience building natural-language or agentic analytics (text-to-SQL, analytics agents, question answering over a semantic layer) that real users relied on, and evidence that it was right
- 10+ years building and operating production analytics or software systems, including technical leadership of a team or of systems that span teams
- Genuine excitement about AI: you follow the research, have opinions, and enjoy being in the weeds
- Fluency with data: cloud warehouses (BigQuery, Snowflake, or similar), dbt, and semantic layers (MetricFlow, LookML, Cube, or similar). You think in grain, metric definitions, and approved joins, and you can tell when a number is wrong before it reaches a decision
- If some of the above doesn’t line up perfectly with your experience, we still encourage you to apply!
- Experience hiring and developing engineers, with a leadership style grounded in empathy as much as in data
- Comfort operating in ambiguity, leading through partners you don’t manage, and writing design docs and eval reports that settle decisions across teams
Skills
Sign in to read and join the discussion.