Builder Developer III

The Builder-3 Profile

You are a systems engineer for non-deterministic systems. You are comfortable saying "no, not yet" to a working demo, and you can say exactly what would change your mind.

1. Core DNA (shared foundation, at lead altitude)

• AI-Native Workflow. You don't just use AI tools fast — you set the squad's bar for what "fast and safe" looks like, and you remove the bottleneck when someone hits one.

• Agentic Architecture. You still build end-to-end agents yourself, and you do. But your leverage comes from designing the state management, tool interfaces and failure semantics that everyone else builds inside.

• Legacy-to-Modern Sequencing. You decidewhich legacy modules get modernized first, based on blast radius, data gravity and how cheaply success can be measured — not on which one looks most fun.

• Measured Reliability. You define what "reliable enough" means for each workflow: acceptance thresholds, regression suites, human-in-the-loop checkpoints, rollback paths. You know that "100% accurate" is a marketing phrase, and that the engineering version is a measured error rate with a defined containment strategy.

2. The Three Things You Own

a) The Guardrail Layer Input and output validation, tool-permission scoping, financial and PII data handling, sandboxed execution, human approval gates on irreversible actions, kill switches, blast-radius containment. You assume the agent will eventually do the wrong thing, and you design so that it costs us a log line instead of a General Ledger.

b) The Eval Framework Golden datasets built from real ERP edge cases. Offline evals in CI as a merge gate, online evals in production. LLM-as-judge calibrated against human labels — and you know when that calibration has drifted. Regression detection when a model version, prompt or retrieval index changes underneath you. You move the squad from "prompt vibes" to a number that can fail a build.

c) The Cost & Performance Envelope Model routing, caching, context budgeting, latency budgets. You track token cost percompleted business transaction, not per call, and you can tell the CFO why a workflow costs what it costs and where the next 40% comes from.

3. Elective Depth (be deep in one, fluent in all six)

• Agentic Architect (Systems) — state management and tool-calling interfaces for long-running business processes across complex ERP schemas.

• UX/Interaction (Product) — generative UIs and browser-use agents; designing where the human stays in the loop and where they don't need to be.

• Data & Context (Data) — RAG and knowledge-graph architecture: vector stores, chunking strategy, retrieval evals, and the relationship between a Purchase Order and a GL entry.

• Reliability & Eval (QA) — programmatic prompt optimization (DSPy), red-teaming agents, adversarial testing of financial logic.

• Infra & Agent-Ops (DevOps) — the factory where agents live: tracing, logging and monitoring the "thoughts" of the AI; deployment and rollback of prompts and models as versioned artifacts.

• Domain Logic Translation (Business) — turning 50 pages of accounting regulation into structured system prompts and agentic constraints.

What You'll Actually Do

• Set the definition of shippable. Write the standard for when an agent can touch production data, and enforce it in code, not in a wiki page.

• Build the eval harness and wire it into CI as a gate that can and does block merges.

• Design the guardrails around regulated and financial logic, including the human approval and rollback paths.

• Run design reviews. Kill bad agent architectures in week one instead of month four. Be specific about why.

• Own agent incidents. Pull the traces, find the root cause, and add the eval that makes that failure impossible to reintroduce.

• Make build-vs-buy calls on frameworks, vendors and models — and revisit them when the ground shifts, which it will.

• Grow Builder 1s and 2s through code review, design review and pairing. Not through ceremony.

What Good Looks Like

(Targets to be agreed with the hiring manager — these are illustrative, not fixed.)

• 90 days: Eval harness running in CI on at least two live agentic workflows, with a published reliability baseline. A written, enforced guardrail standard.

• 6 months: Every production agent has a regression suite, a cost-per-transaction number, and a rollback path. Agent-related incidents are down and post-incident evals exist for each.

• 12 months: The squad ships agentic modules on a predictable cadence without you in the critical path of each one.

The Stack

Python, SQL (Postgres/SQL Server), TypeScript, OpenAI/Anthropic, pgvector/Qdrant, LangChain/LangGraph, DSPy, dbt, Apache Hop/Airflow, Docker, Weights & Biases, OpenTelemetry — plus the orchestration layer you'll live in: eval and tracing platforms (LangSmith / Langfuse / Braintrust or equivalent), policy and guardrail tooling, model routing and caching. And any other tool that helps us achieve greatness.

Who You Are

• You've been burned. You have shipped an agentic system to real users, watched it fail in a way you didn't predict, and changed your architecture because of it. This is the single strongest signal for this level.

• Rigor over hype. You love new models and you still ask for the eval numbers first. You know the difference between a cool demo and a production-grade agent handling financial data.

• Agent-first thinker. You think in capabilities and failure modes, not functions.

• The manual-task hater. If you do a task twice, you've already started drafting the workflow that kills it.

• Technically fearless. A 500-table undocumented legacy schema is a puzzle, not a threat.

• A multiplier. Your best week is one where the squad shipped three things and you wrote one design doc and two brutal, useful reviews.

Why Join Us

• The Final Boss of complexity. ERPs are the hardest problem in software for AI. Automate a General Ledger and you've mastered the game.

• Agentic autonomy. We don't do tickets. We do missions. You choose the models, frameworks and architectures you believe will win.

• No red tape. If the agent works and the evals pass, we ship. No Change Management Committee — a "Does It Work?" policy. Note thatyou are the person who defines what "works" means.

• The Builder Tribe. You will not be the only AI person in the room. You'll lead a squad of people as obsessed with moving 10x faster as you are.