Founding Engineer

Buoyant AI; San Francisco (CA); Hybrid; USD 100,000 - 300,000; Full time

Job summary

Buoyant AI is building the AI layer for engineering consulting firms — AEC, water, transportation, and environmental work. We seek a founding engineer in San Francisco (Hybrid) to ship end-to-end agent-driven workflows, starting with proposal generation inside Microsoft Word.

You will own the generation pipeline, extend the knowledge base, and build evaluation tooling. You’ll help extend one agent into a platform, collaborate with the co-founders, and influence architecture while interfacing

Qualifications

• Strong TypeScript — the entire codebase is TypeScript and you design with strict typing.
• Experience building reliable systems on top of language models and prompt engineering.
• Familiarity with multi-stage LLM pipelines, RAG, evaluation, and deployment in production.

Responsibilities

• Own the generation pipeline end-to-end from prompts to output in Word.
• Extend knowledge base with better search, parsing, and retrieval.
• Build evaluation infrastructure to measure accuracy, voice match, and coherence.
• Make architectural calls shaping the codebase for years ahead.
• Talk to customers, incorporate feedback, and ship features weekly.
• Lead extension of a single agent into a broader platform across domains.

Skills

TypeScript

React

LLM app development

PostgreSQL

Shipping oriented

Tools

Word add-in development

Job description

Engineering HybridFull-time

Founding Engineer

Full-stack product engineer who helps us build the AI platform for engineering consulting firms — agents that own entire professional workflows, starting with proposals.

Location: San Francisco (Hybrid)

Type: Full-time

Compensation: $100,000 – $300,000 + equity

Team: Reports to the co-founders

Who we are

Buoyant is building the AI layer for engineering consulting firms — AEC, water, transportation, environmental. Our bet is that AI can own the dozen knowledge-heavy workflows these firms run every week: proposals, project kickoff, QA/QC, compliance filings, client communications. Anything that isn't the engineering itself is work an agent should do.

Today's product is an AI agent that drafts proposals. It reads the RFP, researches the firm's past work, assembles a team, builds a strategy, and writes — grounded in the firm's real project history, in their voice. It lives inside Microsoft Word because that's where proposal teams already work. It's in production with customers who are winning work with it.

We're backed by $1M+ from Silicon Valley investors focused on AEC and climate. Proposals are the beachhead; what comes next is every other workflow these firms run. The founding engineer is joining at the point where the first agent has proven out and we're about to extend horizontally.

What you'll do

Day to day

Ship features customers are asking for this week

The roadmap is driven by what proposal managers need to win their next RFP — not what we think they might want next quarter.

Own the generation pipeline end-to-end

Better prompts, faster runs, lower cost, cleaner eval. You'll work across the full multi-stage LLM pipeline — content resolution, strategy, writing, compliance review, revision.

Extend the knowledge base

Better search, better parsing, multi-modal support. Vector search, hybrid retrieval, document parsing — all open problems with direct customer impact.

Work inside Microsoft Word

The primary authoring surface is a Word add-in. You'll build task panes, dialogs, and document manipulations that feel native to users who live in Word.

Talk to customers

Engineering firm principals are accessible and opinionated. You'll join calls, run demos, and turn their feedback into product.

Make architectural calls

What to build, what to buy, what to defer. This is a founding role — you'll own decisions that shape the codebase for years.

Who you are

What we're looking for

• Strong TypeScript — the entire codebase is TypeScript. You're comfortable with strict typing, immutable data, and functional patterns.
• React experience — both the web app (Next.js App Router) and Word add-in are React. Server components, hooks, state.
• LLM application development — not ML research, but building reliable systems on top of language models. Prompt engineering, structured extraction, RAG, eval.
• PostgreSQL — schema design, migrations, RLS. pgvector is a bonus.
• A bias toward shipping — you care what users experience, not just what's technically elegant.

Strong advantages

• Office.js or Word add-in experience (rare and valuable).
• Document processing — parsing PDFs, DOCX, and other formats into structured text.
• Supabase / Firebase / BaaS experience — understanding tradeoffs.
• Prior startup experience — comfort with ambiguity, fast iteration, wearing multiple hats.

Not the right role if

• Pure ML/AI researchers who want to train models.
• Frontend-only engineers who won't touch the backend (or vice versa).
• Anyone who needs a large team or established process to be productive.

What you'll build

01

Agents that do senior-level work, end-to-end. A proposal isn't one task — it's scope analysis, team assembly, strategy, writing, compliance review, revision. Design and build the multi-agent system that does all of it, grounded in a firm's real project history, at a quality level where a principal engineer will put their name on the output.

02

Evaluation infrastructure for agentic work. If an agent writes a 40-page proposal, how do you know it's good? Classical NLP metrics fail. LLM-as-judge has its own failure modes. Build that eval infrastructure that measures accuracy, compliance coverage, voice match, and strategic coherence — and catches regressions before they reach customers. Nobody has fully solved this.

03

Context engineering at institutional scale. Firms have decades of documents — thousands of PDFs, resumes, past proposals, specifications. An agent writing about stormwater design in urban California needs the right ten of those, not the thousand closest matches. Better chunking, hybrid retrieval, re-ranking, per-firm learning — this is where the product wins or loses.

04

Extending one agent into a platform. The proposal agent is product one. Next is project kickoff, QA/QC review, compliance filings, client comms. Each is a new domain with its own data model, quality bar, and user workflow — but they share a foundation: the firm's knowledge, their voice, the evals. Build that foundation so new agents ship in weeks, not months.

05

Human–AI collaboration inside the work surface. Most AI products are a chat-box. Ours lives where the work happens — in Microsoft Word, inline with the document, watching edits, learning from feedback. Design the interaction model that turns an agent from a tool you use into a teammate that earns trust over time.

06

The evals flywheel. Every proposal a firm accepts, edits, or rejects is a training signal. Over time we can learn what makes a great proposal for this firm, in this discipline, to this client — and feed that back into the agent. That signal loop, properly built, is the moat. Help us build it.