Senior Software Engineer, AI Research Systems

We build Ursa, a research agent that reads every filing a listed company has published, computes what doesn't line up, and answers a forensic accountant's questions with verbatim quotes and page numbers. Ursa never judges; the expert does. Every expert session is captured as structured data that trains the next version. Our coverage is every US- and Hong Kong-listed company.

We are small and founder-led, and have raised $13M in seed funding; the product is live and institutional subscriptions are open. Most of our code is written with AI coding agents under strict engineering gates; you will be the engineer who owns the system end to end — the one person who answers for what the agents write.

What you will build (in priority order)

1. Ursa itself: a tool-using agent on frontier LLM APIs — retrieval over filings, deterministic calculators, quote verification, lexicon gates, streaming, and the corpus that records every expert session.
2. The expert workbench: a dense, keyboard-first React/TypeScript app on a white page — data grids, citation chips, a working register. Terminal-grade typesetting, no form-filling.
3. The "Said" ledger: an on-prem pipeline on our GPU workstation and storage server that captures executives' earnings calls, investor events and press releases, transcribes and diarizes them, and pairs what was said against what was filed.
4. The public site, the subscriber API, and the cloud that runs them: containerized services behind a load balancer, a Postgres cluster with vector search, object storage with retention locks, scheduled jobs, and CI with a gated deploy chain.
5. Maintenance from day one: you keep what you build running, including the on-prem hardware.

What we need

• 5+ years shipping production software; comfortable owning backend, frontend and infrastructure alone.
• Python and TypeScript/React; Postgres (schema design, roles, migrations); AWS (ECS, Aurora, S3, IAM, EventBridge) or close equivalents.
• Hands-on with LLM applications: tool use, retrieval, evaluation, cost control. You can explain why an agent must never invent a page number and how you would prove it doesn't.
• Some hardware comfort: Linux on a workstation, Docker, a NAS, a GPU, a home-lab mindset. You can set up a storage pool and a WSL2 pipeline without a ticket.
• Evidence-grade habits: hashing, append-only stores, audit trails, least-privilege credentials.
• English working proficiency; Chinese (Mandarin or Cantonese) is a plus — half our filings are bilingual.

Nice to have

• Financial statements literacy, or the patience to learn what receivable days and a KAM are.
• Speech-to-text pipelines, diarization, multilingual embeddings.
• Experience with data-provenance or litigation-hold systems.

How we work

• Facts over opinions, in the product and in the team. Blunt reviews, written decisions, no meetings that could have been a diff.
• Everything deploys through gates; gate changes are separate commits reviewed by a second pair of eyes.
• You will work directly with the founder and with our forensic experts.

Compensation and work authorization

Base salary plus stock options, set by experience and discussed in the first conversation. Candidates must be authorized to work in the United States; visa sponsorship may be considered case by case for exceptional candidates.

How to apply

Email [email protected] with the subject line "Senior Software Engineer — [Your Name]", a short note, and links to work you own end to end (a repo, a system, a write-up). Applying through LinkedIn also works; the note and the links are what we read first. Shortlisted candidates get a paid, time-boxed build task: a small agent that answers questions about one annual report with page-cited quotes and refuses to speculate.