Founding AI/ML Engineer

A $200M solar project gets financed on the strength of a few thousand PDFs: ground leases, interconnection agreements, offtake contracts, permits, title. Analysts read them line by line, by the hour. The documents arrive out of order, in the wrong formats and scanned sideways. The term that kills the deal is on page 74 of a second amendment to an amendment. That review is the bottleneck between capital and megawatts on the grid.

BuildQ accelerates those projects. Our platform has processed 200,000+ documents across 1,100+ projects for the developers, IPPs and funds building multi-billion-dollar solar, wind and storage portfolios.

Most AI companies sell you a tool and wish you luck. We’ve moved to AI-native services. Our customers don’t just use BuildQ to run diligence; they order diligence from BuildQ. An agent builds the deal’s document set, extracts and cites the terms, drafts the site control and interconnection deep-dives, and assembles the IC memo. A human expert signs it. The customer gets a finished work product ahead of their deadline, with our name on the line.

Our customers hand us their most confidential material: unannounced acquisitions, pricing, landowner and counterparty details, deals that haven’t closed. Protecting it isn’t a compliance checkbox. It’s the reason they trust us with it at all. Every system you build treats customer data as strictly isolated, least-privilege and auditable from day one.

And we don’t get to be wrong. A hallucinated term in an IC memo isn’t a bug report, it’s a bad investment decision. Every claim we make must trace to a source page and span. When the evidence isn’t there, the system says so and routes the question to an expert. It never guesses. We’ll ship a “not found” before we ship a confident wrong answer.

We decide with facts, not vibes. Every meaningful change is an experiment: a hypothesis, a metric, a result. “It looks better” doesn’t ship. “Recall on termination rights is up 3 points with no false-positive regression” does.

The details are the job. Which document type regressed? Which clause produced the false positive? Does the gain hold on scanned amendments, or only on clean PDFs? In our world, a one-point bump in false-positive rate means real deals with wrong answers, so we notice it, explain it and fix it before it reaches a customer.

You’ll join our current two-person engineering team as the founding AI/ML engineer. You’ll work closely with our Founding Product Engineer. You own the intelligence: document understanding, retrieval, extraction, predictive risk and benchmarking models, the agent architecture, and the evals that decide whether a change reaches a customer.

The bar isn’t a good benchmark score. A fund underwrites a real transaction on what the system produces. If the extraction is wrong, a fund makes a bad decision. You’ll build systems that are measurably right, that know when they aren’t, and that say so.

• Document understanding on hard inputs. Scanned, rotated, handwritten, redlined and multi-amendment documents. Work out which agreement actually governs, which amendment supersedes which, and what the effective term is today.
• Retrieval over messy deal data. Thousands of documents per deal, with heavy duplication, near-duplicates, and the critical term buried in an exhibit.
• Extraction with citations. Structured terms such as rent escalators, termination rights, interconnection milestones and assignment clauses, each tied to a page and a span an expert can verify in seconds.
• No hallucinations, by construction. Grounding, citation verification, claim-support checks and calibrated abstention. Build systems where an unsupported answer can’t reach a customer. Choose models and providers with data privacy in mind: data retention, redaction, and what leaves our environment.
• Predictive insights. Go beyond what the documents say to what they mean for the project. “Your project is at risk because the interconnection milestone slips past the PPA’s commercial operation deadline.” Turn extracted terms into forward-looking risk signals that customers act on before it’s too late.
• Benchmarking across the industry. We see 1,100+ projects’ worth of leases, interconnection agreements and offtake contracts. Build the benchmarks that tell a customer how their terms, timelines and risks compare with the market, and where they’re an outlier.
• Tabular and decision models. Structured deal data is a modeling surface. Build tabular models for risk and outcome prediction. Use calibrated decision models like TypeSafe’s Jev for routing, scoring and recommending what to do next, not just flagging what’s wrong. Know when each beats an LLM.
• Agent and harness architecture. The agents that assemble document sets, draft deep-dives and write IC memos. Decide what should be an agent, what should stay deterministic, and where confidence should route work to a human.
• Owning the evals. Curate golden sets from real diligence packages. Build verifiers that catch actual regressions instead of giving a green checkmark on the wrong thing. Own the gate for every model and prompt change.
• Learning from experts. Every reviewer correction is labeled data. Close the loop.
• Choosing the right lever. Prompt, retrieval, schema, fine-tune, or just a heuristic, and knowing when ML is the wrong answer for V1.
• Driving down expert hours. Find the hours of review we can safely automate, prove it with evals, and take them.

Rails, React, Postgres, AWS, and Hatchet for agent orchestration. We use Claude, OpenAI and Gemini, with an in-house eval harness. Our own coding agents are harness-agnostic and currently run on Claude Code. You’ll help shape where our ML stack goes next, including a move toward Python for pipelines and evals.

• 2–3+ years shipping ML or LLM systems to production. You’ve handled edge cases, monitored performance and iterated on real user feedback
• Background in NLP or document understanding: RAG, information extraction, OCR or layout models, agentic architectures, fine-tuning
• Experience building tabular or predictive models (gradient boosting, survival or time-to-event, causal or decision models) on sparse, messy real-world data
• You’ve built ML systems where wrong answers had real consequences, and you know how to measure, bound and prevent them, not just lower the average error rate
• Rigorous about evals and experimentation. You catch a meaningful FPR bump or recall drop on a single document type, know whether it’s real or noise, and can systematically debug why. You design measurements you’d bet a deal on
• Strong ML fundamentals and hands-on fluency with modern frameworks and model APIs
• Solid software engineering. Your pipelines are production code, not notebooks

• You’d rather ship three iterations in a month than perfect one model in a quarter
• You’re comfortable talking directly with customers and experts to understand why a model failed and what “good” looks like
• You’re pragmatic: “heuristics for V1, ML in V2”
• You want your work to change how people do their job, not just move a backend metric
• Bonus: you’ve worked in fintech, legal tech, energy, healthcare, or any domain where users stake their reputation on the output

• Meaningful equity at a seed stage company
• 401(k) with a match. Yes, a startup offering a match.
• Full health coverage for you and your family
• Remote (US)
• Whatever you need to get set up well at home
• All the AI tools and tokens you need

Email careers@buildq.ai with your resume and a short note about something you shipped that you’re proud of.