Agent Engineer

Rifa AI; San Francisco (CA); On-site; USD 120,000 - 180,000; Full time

Job summary

Rifa AI, based in San Francisco, builds an AI agents platform for contact centers in regulated industries. We create voice and chat agents that operate within client rules and integrate with CRMs, with automated testing and post-analysis to ensure compliance.

As an Agent Engineer, you own a client's voice or chat agent in production, including the procedure it follows, its instructions, and the integration tests.

Qualifications

• 2 to 5 years building and running production systems.
• You write Python another person can read, review others' work thoughtfully, and debug by forming a theory and testing it, not by changing things until they work.
• You've worked with async code and a relational database.
• Comfort working directly with customers to understand their needs and solve real-world problems: take vague feedback, ask the right questions, leave with a scoped change.
• Strong written communication. Agent instructions, client explanations, and incident writeups are all writing, and an agent is only as precise as the instructions behind it.

Responsibilities

• Own a client delivery end to end. Take requirements from first conversation through pilot, production, and continuous iteration as procedures change, volumes grow, and models improve.
• Engineer the agent's behavior. Write and maintain the instruction sets that determine what the agent says and when.
• Build the evaluation gate. Automated tests that replay past conversations and check the agent follows the procedure.
• Debug live conversations. Look up what the system knew at each moment, read the exact instructions the model was given, and work out why it responded the way it did.
• Work directly with clients. Sit in on calls with US enterprise clients, own the technical conversation, and watch your changes move their business metrics.

Skills

Python

Async code

On-call experience

Relational database

Customer interaction

Written communication

Tools

FastAPI

PostgreSQL

Temporal

WebSockets

STT/TTS providers

OpenAI

Job description

About us

Rifa AI is building the AI agents platform for contact centers in regulated industries.

Enterprises in these industries want AI agents handling their customer operations and mostly can't deploy them. It's not a model problem. Horizontal platforms lack governance, release processes, and change management, and in a domain where every call can be reviewed by a regulator, that's disqualifying. Building an AI agent has never been easier. Deploying one an enterprise can trust has never been harder. That harder problem is the one we work on.

Our platform turns a company's written procedures into AI agents: voicebots that hold real-time conversations, take actions in the client's CRM, and stay within the limits the client has set. Every release is gated by an automated testing suite, and every conversation feeds a post-analysis platform. We're live in production today, handling debt collection calls for US financial services clients.

The engineering convictions behind it: evaluation methodology, not model capability, is the bottleneck. Every rule and decision trace becomes part of a company's context graph. Observability goes beyond logging. And our Agent Studio lets engineers, non-engineers, and auditors collaborate to build and improve AI agents with the right guardrails to achieve the expected business outcomes.

Rifa was founded by Sameer Fulzele (IIT Bombay). We're a small team of exceptionally capable, passionate engineers with paying enterprise clients and growing revenue, backed by Seaborne Capital, a founder-first firm of exceptional industry operators that works closely with its founders, and by angel investors who are veterans of enterprise software and operators in the accounts receivable industry.

What you'll do

An Agent Engineer owns a client's voice or chat agent in production. Not a component of it, the whole thing: the procedure it follows, the instructions that govern how it speaks, the code that connects it to the client's systems, and the tests that prove it does what the documentation says. When a client says "the bot offered a payment plan below our minimum," you're the person who works out why, fixes it, and explains the fix to the client. You'll ship changes that speak to real callers in your first two weeks.

• Own a client delivery end to end. Take requirements from first conversation through pilot, production, and continuous iteration as procedures change, volumes grow, and models improve.

• Engineer the agent's behavior. Write and maintain the instruction sets that determine what the agent says and when. Instructions are versioned, tested, and reviewed like code, because a wrong word in a disclosure is a compliance incident, not a UX bug.

• Build the evaluation gate. Automated tests that replay past conversations and check the agent follows the procedure. Nothing ships without passing them.

• Debug live conversations. Look up what the system knew at each moment, read the exact instructions the model was given, and work out why it responded the way it did.

• Work directly with clients. Sit in on calls with US enterprise clients, own the technical conversation, and watch your changes move their business metrics. Not many engineering roles put you this close to the people using what you build.

Example projects

Recent work by engineers on this team, all on live debt collection agents:

• Build an intent identification layer for a voice agent, so every response is grounded in what the caller actually asked rather than what the model assumes, sharply reducing hallucinations on live calls

• Extend the negotiation flow so the agent offers payment plans only within the limits the client has set, with guardrails that make out-of-bounds offers impossible rather than just unlikely

• Grow the eval suite that replays real collection calls and verifies every legally required disclosure was delivered, word for word, before any release ships

What we work with

Python with FastAPI, PostgreSQL, Temporal for background workflows, WebSockets holding real-time conversations open, and STT and TTS providers on the speech side, with LLMs via OpenAI and similar underneath.

Everything runs on Kubernetes with ArgoCD-driven GitOps deployments, SigNoz for observability, DeepEval driving the eval suite our CI runs before anything reaches production, Ory and OpenFGA for identity and access management, and a lot of beautifully built internal agent architecture underneath.

What you’ll bring

• 2 to 5 years building and running production systems. You've been on call for something you built and debugged it under pressure.

• You write Python another person can read, review others' work thoughtfully, and debug by forming a theory and testing it, not by changing things until they work. Reading code you didn't write and working out what it does is the single most important skill in the role.

• You've worked with async code and a relational database.

• Comfort working directly with customers to understand their needs and solve real-world problems: take vague feedback, ask the right questions, leave with a scoped change.

• Strong written communication. Agent instructions, client explanations, and incident writeups are all writing, and an agent is only as precise as the instructions behind it.

Even better

• LLM systems in production: eval frameworks, agent tooling, RAG pipelines, structured prompting

• Conversational AI experience, voice or chat: dialogue design, IVR systems, chatbots, speech interfaces

• A regulated industry: finance, healthcare, insurance, collections

• Real-time or telephony systems: WebSockets, streaming audio, Twilio or similar

• Founder or founding engineer experience

Explicitly not required: prior "AI engineering" as a job title, or a computer science degree. Careful engineers who own outcomes pick this up fast.

How we work

Small teams per client, with real ownership. You'll have peers to pair with and review your work, and you'll review theirs. We review all code, we write things down, and we'd rather hear an honest "I don't know yet" than a confident wrong answer.

Our values

• Trust: We do the right thing, especially when nobody is checking. Clients hand us regulated conversations with their own customers, and we earn that every day.

• Transparency: We write things down, share the real numbers, and say "I don't know yet" out loud, with each other and with clients.

• Technically best solution: We choose what's right, not what's easiest or trendiest. When we notice we got it wrong, we fix it.

• Decisiveness: We decide quickly with the information we have, commit, and correct course fast when reality disagrees.

• Simplicity: We keep systems, processes, and words simple. Complexity is a cost we pay only when it clearly buys something.