Software Engineer
About Conduit
Conduit is an AI agent platform purpose-built for hospitality. Hotels, vacation rentals, and resorts deploy our agents across guest-facing communication and internal operations, running millions of workflows a day.
Hospitality is about 10% of global GDP and it turns on customer experience. Resolving a guest request immediately is most of that experience, and resolving one means moving across several tools and between several people. That coordination is the work, and it is what our agents do.
Role: Software Engineer
You'll own the core intelligence layer of the product: the system that reads a customer message, decides what to do, retrieves the right context, takes action, and explains its reasoning. This isn't a research role. You're building production AI systems that handle real conversations across SMS, email, WhatsApp, voice, and web chat, in real time, with real consequences when they get it wrong.
You'll work directly with Punn (CTO) and a small engineering team where what you ship today changes how thousands of conversations go tomorrow.
Why this role is different
You're solving a data problem, not a model problem. LLMs don't really hallucinate anymore. They faithfully ground answers in whatever context they're given. The hard problem is making sure that context is correct, complete, and current when enterprise data is none of those things. You're building the retrieval, ranking, and validation systems that make AI reliable despite messy real-world data.
You're building AI that does things, not just says things. Our agents don't describe how to process a refund. They process it. You'll design the tool-use and function-calling layer that lets AI agents execute multi-step workflows: modify reservations, dispatch maintenance, send confirmations, close loops. The gap between "AI that talks" and "AI that acts" is where you live.
You're making AI auditable. Every other platform in this space is a black box. You're building the decision trace: what the AI retrieved, what it considered, what it ignored, where confidence dropped, and why it chose what it chose. This is what lets CX teams actually trust and improve the system instead of praying it works.
You're building for non-technical operators. The people training and managing your AI system aren't engineers. They're support leads who've handled 10,000 conversations. Your job is to make the AI's behavior legible, testable, and adjustable by someone who's never written a line of code.
What the work looks like
Response generation pipeline. The router-orchestrator-synthesizer architecture that takes an incoming message, classifies intent, assembles context, selects tools, generates a response, and decides whether to send it or escalate. Latency matters. Accuracy matters more.
Knowledge retrieval. Hierarchical knowledge base with scope-based queries. The AI needs to pull the right policy for the right property for the right situation, not just the closest embedding match. You'll build the retrieval layer that handles contradictory docs, stale articles, and edge cases that only exist in one rep's head.
Tool orchestration. Defining, exposing, and managing the tools available to AI agents. Connecting workflows-as-tools so the AI can take real actions. Handling tool failures gracefully when a third-party API is down mid-conversation.
Evaluation and testing. Building the eval framework that lets the team measure AI quality against real conversations, catch regressions before they hit customers, and give CX teams confidence to expand automation. If you can't measure it, you can't improve it.
AI observability. Structured traces for every AI decision. Source attribution. Confidence scoring. The infrastructure that powers the "cockpit" view where operators can see exactly why the AI said what it said and fix the root cause in minutes.
Voice AI. Real-time voice conversations with transcription, intent detection, and the same tool-use capabilities as text channels. Different latency constraints, same reliability bar.
Speed and cost optimization. Model routing, caching strategies, prompt optimization. Making the system faster and cheaper without sacrificing quality. Every 100ms matters when someone is waiting for a reply.
Who you are
You've built production AI/ML systems that serve real users, not just notebooks and demos.
You think about retrieval as seriously as generation. You know the answer is only as good as the context.
You're pragmatic about models. You pick the right model for the job, not the most impressive one. Sometimes that's a frontier model. Sometimes it's a fine-tuned small model. Sometimes it's a regex.
You care about evaluation. You don't ship a prompt change without knowing what it does to accuracy across 1,000 real conversations.
You want to own the full loop: retrieval, generation, tool use, tracing, evaluation, and the feedback cycle that makes it all better over time.
You want a small team where your work directly shapes the product and the AI's behavior changes the day you push it.
Compensation & Benefits
$130-220K base
Ownership of the core intelligence layer of the product
Day-1 influence on AI architecture and strategy
Direct line to the CTO and the customer
Interview Process
First Call (20 min). Quick conversation with the founders. We'll talk about what you've built, how you think about AI systems in production, and whether the comp range works for both sides. No prep needed.
Technical Interview (45 min). An AI systems problem: debugging a retrieval pipeline that's returning wrong context, designing a tool-use architecture for a multi-step workflow, evaluating why an AI response went sideways on a real conversation. No leetcode.
Optional: Agent Coding Session (5 min video). Record a short screen capture of you working through a problem with an AI coding agent (Cursor, Claude Code, Copilot, whatever you use). Not required, but it tells us a lot about how you work.
On-site Work Day (1 day, San Francisco). Spend a day with the team. We'll scope a small AI feature together, you'll build and ship it. Travel and time are compensated. Remote option available if SF doesn't work.
References & Offer.
If you read this and immediately started thinking about how you'd structure the retrieval layer, we should talk.