Founding AI Engineer

Specter; New York (NY); On-site; USD 180,000 - 240,000; Full time

Job summary

Specter is seeking a lead for autonomous accounting to own the agent loop, context systems, and production-ready infrastructure that can act on real general ledgers without breaking them. You’ll tackle hard problems like giving an agent a defensible view of a company’s books, measuring decisions, and making the system better every week.

You will design architectures for accounting workflows, build the knowledge capture of customer domains, and develop evals with customer-specific tests and

Qualifications

• Experience with modern LLMs and agent-based systems.
• Ability to model complex accounting workflows.
• Proven track record shipping production systems.
• Comfort navigating ambiguity in new architectures.

Responsibilities

• Design architectures for accounting workflows where decisions cost dollars.
• Build systems to capture customer domain knowledge and make it retrievable.
• Develop evals predicting production behavior with fixtures and criteria.
• Turn research into constrained experiments with gates and audits.
• Set technical direction for how agents reason, retrieve, verify, and improve.

Skills

LLM systems

Distributed systems

Data modeling

Latency optimization

Agent loops

Cost awareness

Tools

Evaluation frameworks

Retrieval systems

Audit trails

Open-source agent libs

Job description

The role

Autonomous accounting is not a wrapper problem. It's an agent problem, a context problem, and an evaluation problem all at once, and the team that solves it will be the one that treats those as engineering work, not prompt engineering work. This role is for someone who wants to own that.

You’ll work on the agent loop, the context systems that feed it, and the infrastructure that lets our agents act on real general ledgers without breaking them. The hardest technical problems at Specter live here: how do you give an agent enough understanding of a specific company's books to make a defensible decision, how do you measure whether that decision was actually right, and how do you make the whole thing better every week.

What you'll do

• Design agent architectures for accounting workflows where the cost of a wrong decision is measured in dollars, not impressions. Reconciliations, journal entries, AP coding, revenue recognition, accruals, intercompany.
• Build the systems that capture each customer's domain knowledge - chart of accounts, recurring patterns, approval rules, exceptions - and make it retrievable in the right shape at the right moment.
• Develop evals that predict production behavior using customer-specific fixtures, acceptance criteria, and controller review outcomes alongside generic benchmarks.
• Turn promising AI research into constrained production experiments behind eval gates, audit trails, and customer-specific acceptance tests. The feedback loop from idea to customer is short, and you control most of it.
• Set the technical direction for how Specter's agents reason, retrieve, verify, and improve over time. Decisions made in this role compound.

What we look for

• Deep familiarity with modern LLM systems: tool use, retrieval, fine-tuning, structured output, agent loops. You don't need a PhD; you need taste and the ability to tell a working idea from a marketing idea.
• Strong fundamentals. You can reason about distributed systems, data modeling, latency, and cost, not just prompts.
• A track record of shipping. Research papers are great; production systems other people depend on are better.
• Comfort with ambiguity. The right architecture for autonomous accounting is not in a blog post yet, and the person we hire here is part of how it gets figured out.

Nice to have

• Background in finance, accounting, audit, or any domain where correctness is non-negotiable and "looks plausible" is a failure mode.
• Open-source contributions to agent, retrieval, or eval frameworks.
• Experience designing systems that learn from human feedback in production.