We're hiring two founding engineers.

Pipa — the product layer

The Pipa toolkits surface is the part of the product the HR person uses day to day — the workflows that take them from a first hiring conversation, through onboarding, and into the harder moments of the employment relationship. We've shipped it to our first design partners. It works, customers like it, and it's not yet ready for paying users.

That's the role. You'd own the production-readiness layer end to end: auth, multi-tenancy, security, observability, deployment pipeline, cost discipline, recovery. Roughly three to four months of focused work to ship a paying-customer-ready product. After that, the work shifts: scale, performance, integrations, the next AI surfaces. You'd be the senior engineering voice for everything the user touches.

You'd pair with Lee on product calls and with Claude on most code. This is the operating rhythm at Pipa — not a preference. Most of the codebase is built daily with Claude. If "I'd be conducting AI output" reads as exciting, this is the right rhythm for you; if it reads as babysitting, it isn't.

Must-haves

• Senior full-stack JavaScript / TypeScript. You'll be productive day one in a hand-built codebase without retraining.
• Production engineering instincts. Monitoring, logging, secrets management, rate limiting, incident response — you've shipped these in real products and lived through the 3am page that taught you why they matter.
• Security-minded. Auth, multi-tenancy, and GDPR-aware data architecture are your first three deliverables. You think about auth like a person who's been the one to write the post-mortem.
• Comfortable in a prototype-to-production codebase. Pipa is hand-built and a bit scrappy. Someone who needs a clean React + TypeScript + Vite + tRPC monorepo to function isn't the fit. You can read what's there, respect the decisions that were already made, and upgrade thoughtfully where it matters.
• Product-minded, willing to push back. Not an executor. You argue about UX and tell Lee when she's wrong.
• Low-ego, direct written communication. Our rhythm is async, written, blunt-but-warm. You can say "I don't know" and disagree without softening.
• Fluent with AI-assisted development. You brief Claude to elicit useful work, read its output critically, push back when it's wrong, and integrate AI output without copy-paste sprawl. This is how the velocity exists.
• Personal runway and the right shape of leap. Co-founder-equivalent role at a pre-funded company. If you have unemployment, savings, or partner-income runway and you've been wanting to bet on something, this is exactly that shape.

Strong preferences

• Hebrew fluency and Israeli market familiarity. Pipa is Hebrew-first.
• HR-tech, B2B SaaS, or people-operations domain familiarity. Changes the speed of every UX-shaped product call.
• Has owned a multi-tenant app from auth to billing. You know what "users sometimes wear three hats and switch organizations" actually means in practice, not just in theory.

Nice-to-haves

• Prior founding engineer or founder experience.
• Realtime database experience, or pragmatism about when migration makes sense versus when it doesn't.
• AI / LLM product experience as a developer, not just a user.
• Design sensibility — Pipa cares about craft.
• Network in the Israeli operator-tools or HR-tech space.

This probably isn't for you if

• "Let's rewrite this in Next.js / Remix / Bun" is your first instinct.
• You don't believe in the product but the equity narrative sounds interesting.
• You treat AI as a magic wand and won't own the boring 200 production things.
• You want to be the architect, not the operator.
• You need heavy structure — no PM, no Jira, no daily standup framework.
• You've only ever worked on greenfield projects and find inheriting code uncomfortable.

Alma — the deterministic AI side

Alma is a deterministic emotional-development AI built in Python — a state machine, signal detection, lexical wall, and LLM-as-extractor pipeline. She encodes a 256-page behavioral spec into roughly 23,000 lines of code with 200+ regression tests. Lee is the architect and the spec keeper. You'd be the second engineer, working specifically on Alma.

The work isn't model training or research. It's translating dense behavioral spec into deterministic dispatch, hardening the system against LLM drift with lexicon walls and eval-gated classifiers, and chasing live-conversation bugs into spec-cited fixes with multi-layer regression tests. You'd pair with Lee on architectural calls — dispatch ordering, signal taxonomy, spec ambiguity — and with Claude on most code. Alma is 85% Claude-written today; that's not a preference, that's the rhythm.

Must-haves

• Senior Python systems engineering. You can navigate a 23K-line codebase, hold dispatch logic in your head, and design test scaffolding that pins behavior across signal detection, dispatch, end-to-end, and negative controls. Production-grade Python, not "I can write a script."
• Worked with LLMs as components, not as the brain. You've shipped at least one production system where the LLM is part of a deterministic pipeline. You understand operator-level prompt engineering: temperature 0, pinned model snapshots, structured outputs, safe-default-on-failure. You treat the LLM as a stochastic component that requires guardrails.
• Test-driven discipline by reflex. Every bug that escapes Alma to production becomes a new regression test. You write tests as part of the fix, not after. You design multi-layer pins. Green CI is non-negotiable.
• Spec patience. The spec is dense — many sections, cross-referenced rules, NEVER / ALLOWED lexicons, compound gates. You sit with prose, translate it to code, cite the spec section in comments, and resist the urge to "improve" what you don't yet understand.
• Low-ego, direct written communication. You can say "I don't know" when you don't, disagree with Lee in writing without softening, and read a "this didn't hold in live conversation" reply without a defensive crouch.
• Fluent with AI-assisted development. You know how to brief Claude to elicit useful work, read AI-generated code critically, push back when the AI is wrong, and integrate AI output without copy-paste sprawl. Alma is 85% Claude-written today; this isn't optional.
• Personal runway and the right shape of leap. Same as the Pipa role. Co-founder-equivalent, mostly equity, real bet.

Strong preferences

• Hebrew fluency and Israeli market familiarity. Alma has Hebrew-first edge cases. Cultural and timezone alignment accelerates onboarding meaningfully.
• Background or interest in psychology, coaching, or behavioral science. Alma encodes emotional-development theory. The diagnostic conversation "this output is wrong" is half-semantic, half-code. Domain familiarity changes how fast you can read a live-conversation screenshot.
• Deterministic or rule-based NLU experience. Pre-LLM intent detection, slot-filling, state-machine dialogue systems. You've seen what breaks when those get replaced with pure-LLM and know why Alma is built the way it is.

Nice-to-haves

• Eval-driven development background. JSONL corpora, tiered accuracy thresholds, variance testing, CI gates on regression.
• Small-company instincts. Sub-10-person engineering teams, no PM, no QA, no dedicated infra. Comfortable being all of those in rotation.
• Production observability discipline. "What would I want to see in the logs when this breaks at 3am" as a default before shipping.

This probably isn't for you if

• Your portfolio is heavy on model training, fine-tuning, or research. Alma doesn't train.
• Reading 80 lines of "why this dispatch order" makes you want to refactor instead of reading. Half the codebase's value is the comments and the spec citations.
• You're a frontend specialist looking for generalist work. The Pipa toolkits side is covered.
• You need heavy structure. No PM, no Jira, no daily standup framework — the structure is the spec, the tests, the live-conversation screenshots, and Lee's strategic calls.
• Your instinct on every problem is "let's just let the LLM cook." The architectural commitment is that Alma is NOT generative — she's a deterministic extractor, dispatcher, and safety-walled responder. You'd fight the architecture every day.