Applied Engineer; Tribal

Applied Engineer (Tribal Knowledge)

Design and own the agent system that compiles an organization’s tribal knowledge

London / SF Full-time Engineering

Pavo is building Enterprise Superintelligence: compounding systems that take ownership of business outcomes and work with humans to deliver them.

We believe that while foundation models are necessary, they are not sufficient. The hard problem is systems intelligence: end-to-end architectures that understand a company’s code, data, and decisions, and improve themselves through experience.

We are assembling a small, senior team of researchers and engineers obsessed with systems-first intelligence. Our current team consists of PhDs and ML engineers from top applied ML and coding agent companies, with a heritage of shipping systems at Spotify, Share Chat, and Sourcegraph scale.

Our team has built impressive momentum with a small group of highly capable engineers and researchers.

As an Applied Engineer at Pavo, you will design and own the agent system that compiles an organization’s tribal knowledge — the body of knowledge that nobody has written down — from its own evidence: source code, structured data, internal documents, and conversations. You’ll work hands-on at the level of agent architecture, pipeline design, and the production surface that turns a generative system into one customers can depend on.

You’ll partner closely with the Applied Scientist on instrumentation — but the system itself is yours. It is the most consequential thing we build, and idea to production is days.

This is a senior, individual-contributor role. Everyone on the team joins as a Member of Technical Staff — with the scope, autonomy, and end-to-end ownership that title implies.

• Hierarchical Agent Architecture:
  Design how a main agent coordinates sub-agents, tools, and skills — deciding where deterministic scaffolding beats prompting and where prompting is the right tool. This decomposition is the single biggest determinant of whether a generative knowledge system is dependable.
• The Synthesis Pipeline:
  Own the multi-stage system that turns heterogeneous private evidence into a verifiable knowledge artifact — with the checkpoints, retries, fallbacks, and structured outputs that production reliability requires.
• Deployment & Rollout:
  Build scheduled regeneration, quality and reliability gates, blue/green rollout, and customer-facing version control of compiled knowledge artifacts. Treat each release with the seriousness it deserves.
• Observability & Debuggability:
  Structured traces of agent runs, replay/debug tooling, cost and latency budgets, and regression detection on releases. A multi-stage agent run must be debuggable by someone who didn’t initiate it.
• Reliability as a First-Class Property:
  Treat run-to-run variance as a defect class equal to incorrectness, and build the engineering practice that reduces it.
• Eval & CI Infrastructure:
  Reproducible, automated evaluation that feeds signal back into the deploy loop — complementary to the experimental harness on the science side.

We are looking for an engineer who has shipped dependable production systems and is hungry to do it at the edge of what current agentic frameworks can do.

• Senior Track Record: 8+ years building and operating production systems, including a stretch where you owned a major system or platform area end-to-end and were the engineer others came to for the hard problems.
• Production Engineering Chops:
  You ship Python systems that other engineers want to extend; you write tests that catch real bugs; you have opinions about structure, dependency boundaries, and what belongs in a library versus an application.
• Hands‑On with an Agentic Framework:
  Depth with at least one — Anthropic tool use, OpenAI Agents SDK, Pydantic AI, Lang Graph, Auto Gen, or equivalent — including its sharp edges, not just its happy path. We have no preference for which; we strongly prefer that you’ve been bitten by one.
• Strong Opinions on Agent Architecture:
  When to use a single-context agent versus decompose into sub‑agents. Where deterministic code beats prompting,…