Job purpose
We’re building a voice agent that calls people who are past due on bills, explains their situation, and negotiates a payment arrangement
–
in real time, without a human on the line. It uses synthetic voice, behavioral science, and LLM-orchestrated decision logic to run conversations that actually work. That’s the product. It’s technical, ethically loaded, and genuinely hard to build well.
This role exists at the intersection of product and engineering, and it leans hard into both. You’ll own the full lifecycle of that product: pre-sales conversations where you help a client understand what’s possible, discovery where you find the real job-to-be-done, and implementation where you build and ship it. You’re not a PM who hands off to engineering. You’re not an engineer who gets handed a spec. You do both, and you’re accountable for whether the product finds its market.
You’ll lead a small delivery team
,
one to two junior and intermediate engineers
,
operating on a parallel track to the core platform. What you learn feeds back into the roadmap. What you ship is the
conversational
product.
The domain matters. You’re building AI that makes real-time decisions with real financial consequences
–
inside compliance guardrails, with a person on the other end of the line. Candidates who have worked in regulated, high-stakes AI application domains will have a meaningful head start. If you’ve survived a client integration that went sideways at 11pm and have opinions about when to throw out what the model generated, read on.
What You'll Do:
LLM orchestration depth
is the floor
You’ve built real agentic pipelines. You understand context window constraints, tool call patterns, failure modes, and what happens when an agent loses the thread mid-call. “I’ve used ChatGPT” is not this. You should be able to explain where your last pipeline broke and why.
Voice AI and telephony familiarity
Hands-on experience with at least one of: Twilio, ElevenLabs, Retell, or Deepgram. You know why real-time audio is technically hard
–
latency, media streaming, session state
–
and you’ve built something that had to work under those constraints.
Full-stack implementation capability
You write async Python with confidence:
asyncio, FastAPI, SQLAlchemy async. That’s the core of this stack and where most of the hard problems live. You’re comfortable in TypeScript/React and can work across the frontend, but you don’t need to be a specialist there. If you’re stronger on the backend and can navigate a Vite/TanStack/Zustand frontend without getting lost, that’s fine.
Client-facing discovery skills
You know how to run a conversation that surfaces the real problem, not just the stated one. You can translate a messy client requirement into a scoped technical approach and explain it back to a non-technical stakeholder without losing the room.
Integration and event-driven architecture
You’ve wired up external APIs, dealt with inconsistent documentation, rate limits, and auth schemes, and made it reliable. Webhooks are
the primary integration pattern for Twilio, ElevenLabs, and Retell. You’ve built webhook receivers, implemented HMAC signature validation, handled replay attacks,
and designed for idempotency. You’ve used Redis as more than a cache: pub/sub, streams, session state at scale.
AI tooling philosophy
You use AI as a drafting partner, not a decision-maker. You reach for it on boilerplate, test scaffolding, and exploration. You own the design. You know when the output needs to be thrown out. A bad sign: you ship what the model generates without reading it.
Multi-tenant systems experience
You’ve built multi-tenant SaaS where tenant data bleed is a real threat, not a hypothetical, and you can describe specifically how you prevented it. Per-tenant credential scoping, namespaced keys, org-scoped access patterns
are
second nature.
Cloud deployment literacy
You’ve shipped to cloud-hosted containers. You can read a GitHub Actions workflow, understand why a build fails in CI but works locally, and debug a deployment without needing someone to translate the logs. Azure experience is a plus; the pattern matters more than the platform.
Agentic SDLC toolkit fluency
We’ve built a custom agentic development toolkit on top of Claude Code and internal tooling. You’ll become expert in it. Not just as a user, but as someone who understands its capabilities well enough to identify where it falls short and push it forward. Existing
AI Coding Agent
experience
is a strong signal
(Claude Code, Codex, OpenCode, etc.).
Performance Expectations
90 days:
First client implementation is in active development with a clear delivery plan and a client aligned on scope. You’re fluent in the agentic SDLC toolkit. You’ve led at least one discovery conversation independently and can articulate where the product fits and where it doesn’t for that client’s context.
12 months:
Three to four clients live on the voice AI product. A repeatable delivery motion exists — not just in your head, but documented and transferable. You’re feeding clear PMF signal back into the core roadmap, and there’s a case for growing the team or doubling down on the product direction.
What You'll Need:
Minimum 3 years experience building SaaS products at scale
B2B experience is an asset
You’re probably not a fit if
You need detailed tickets and a tech lead to make forward progress
You’ve built “integrations” but couldn’t explain the failure mode at the API boundary
You’ve “worked with LLMs” but haven’t built a production agentic pipeline
You’ve never been in a client conversation and don’t want to be
You treat AI-generated code as ground truth