Principal Software Engineer (AI Specialist Track)

As a Principal Software Engineer (AI Specialist Track), you will own the AI architecture across Planet Depos’s product surface and serve as a Principal-level peer on our architecture council. This role exists to build AI systems that do meaningful work on behalf of users – identifying inconsistencies across testimony, surfacing issues an attorney may have missed during a live deposition, and drafting questions for tomorrow’s deposition based on what was filed yesterday – rather than simply creating “chat with a transcript” experiences. 

Today, AI architecture decisions sit within the broader principal engineering team alongside many competing priorities, which doesn’t match the shipping pace this role exists to enable. This seat absorbs that load on the AI side, with full authority to make model-selection / orchestration / retrieval / structured-output decisions independently.

Essential Job Responsibilities:

• Set the agentic-first development pattern for the team — pair-program with AI tooling visibly, mentor without it being a job title, model what “agentic but with eval discipline” looks like in practice. The team upskills toward AI fluency through the architecture patterns you ship, not through certifications.
• Own AI architecture decisions independently — model selection per task class, agent orchestration patterns, RAG retrieval design, structured output, prompt engineering at the platform level. Co-author architecture decision records (ADRs) for AI-touching work; sit on the architecture council as a Principal-tier peer.
• Own product/feature evaluation discipline — per-feature accuracy, jobs-to-be-done success rate, hallucination measurement, the eval gates that gate-keep production launch.
• Ship agentic product surfaces in production — chat-with-case, live-during-deposition intelligence, pre-deposition inconsistency analysis, post-proceeding workflows. Visible client adoption inside 12 months.
• Frame what gets built around jobs-to-be-done — not “chat with X.” Define the specific outcomes the agent has to deliver, then build and measure against those outcomes. Translate client and stakeholder signal into agentic features that do work, with citations, that someone can actually use in a deposition.

Principal Software Engineer (AI Specialist Track) Location:

Fully Remote

Principal Software Engineer Compensation:

$250,000+ base compensation and bonus eligible, commensurate with experience

Benefits:

• Medical
• Dental
• Vision
• Voluntary Term Life Insurance
• Voluntary Whole Life Insurance
• Voluntary Long Term Disability
• PTO
• Paid Holidays
• 401(k)
• Employee Assistance Program (EAP)
• Maternity Leave

Required Qualifications: You must possess all three qualifications listed below. Candidates who possess only one or two qualifications will not be considered for the role. 

1) Experience in shipping agentic systems that do work, not chat, in production, with evaluation discipline 

You’ve built and shipped systems where the agent acts on behalf of a user — not “summarize this transcript,” but “find the inconsistencies between these three depositions and surface the questions that would expose them.” Production means real users, real consequences, real eval gates. The agentic-first thesis — agent does work, doesn’t just chat — is your default, not a stretch you’re aspiring to. 

Qualifying experiences include: 

• An agentic system you shipped where the agent took multi-step action on user behalf, with the eval harness that validated it was production-ready before launch — accuracy floor, known failure modes, the cost model, and the graceful-failure path when the agent gets it wrong 
• A jobs-to-be-done framing you applied to an AI feature: instead of “chat with X,” you defined the specific outcomes the agent had to deliver, then built and measured against those outcomes — and can describe the chat-surface temptation you turned down 

2) Principal-tier judgment on AI architecture — model selection, agent orchestration, RAG, structured output, retrieval 

You've owned major AI architecture decisions yourself and can justify them, as opposed to implementing as designed by others. You understand how to evaluate model selection per task class, agent orchestration patterns (single-agent vs. multi-agent and the associated tradeoffs), and RAG retrieval design (chunking, embedding choices, re-ranking, and hybrid search). You can intuitively evaluate best use cases for structured outputs vs. free-form generation, and frontier-vs.-distilled model selection per task. You can document those decisions clearly and create architectural guidance that enables a broader engineering team to scale AI development beyond a single individual. 

Qualifying experiences include: 

• An AI architecture decision you made — model selection, orchestration pattern, retrieval design — and can describe the alternatives you rejected, why, and what changed in production that proved or disproved your call 
• A team you upskilled toward AI fluency through architecture patterns you shipped, not through training programs or office hours — the mentorship was structural (you set the patterns the team built against), and you can name the engineers who came up through it 

3) Shipped under compliance-sensitive domain stakes — legal, healthcare, finance, regulated tech 

A hallucinated answer where the answer matters is brand damage with real stakes. You’ve worked somewhere wrong-output had consequences — legal, healthcare, financial services, insurance, regulated tech — and you can describe what guardrails you put in, what you measured, what you refused to ship. 

Qualifying experiences include: 

• Built software in a domain where a wrong AI output had legal, medical, or financial consequences — and you can name the specific guardrails (evaluation gates, human-in-the-loop patterns, refusal modes, citation requirements) you put in and how you validated they actually worked in production 
• Has handled real enterprise privileged communications, PII, or PHI in production AI systems — knows the protocols, knows where enterprise-tier model contracts protect you vs. where you have to build your own protection 

EOE M/F/D/V