Original job description
Fully Remote · Occasional travel for team coworking and client visits · Reports to VP of Technology · $120,000–$165,000 base + equity
About Your Role
Softrip is transforming from a traditional SaaS travel technology company into an AI-first platform - not by bolting AI features onto existing software, but by rethinking what the product can do when AI is native to it. Of our existing customer base, we have beta customers ready to build with us, solid discovery work already done, and a clear direction. We need dedicated engineers with external experience doing exactly this to help us deliver - and to keep pushing the boundaries of what’s possible as our customer base grows and our AI capabilities expand.
You’ll ship AI-powered features into both of our products - one serving enterprise travel companies, one serving SMB operators - and build the engineering layer underneath them. You’ll be hired alongside an AI Product Manager as a deliberate two-person AI team. The working model is iterative and close: you’re not receiving specs and building to them. You’re figuring out what to build together, in real time, and shipping it.
The foundation is in place: Langfuse is running for LLM observability and LaunchDarkly is configured so prompts can be updated and pushed to production without a code deploy. You’ll build on top of that foundation and extend it as our capabilities and customer demands grow.
The team is small by design. There’s no dedicated design function, no QA layer, and no DevOps handoff. You own what you build end to end, from initial integration to production reliability. If your instinct is to wait for sign-off before making a call, this won’t fit. If you’ve been the person who figures things out and ships them without a net, keep reading.
The Job
You ship AI features from week one. Not a proof of concept - features in the product that real beta customers interact with. You start focused, learn what the product and customers actually need from AI, and grow the complexity from there.
You build context pipelines. The model isn’t the hard part - assembling the right product data at the right moment is. Getting customer and operational data assembled efficiently, within context limits, and in a format the model can reason over is real engineering work. You’ll solve it per feature and get better at it with every iteration.
You build and own the eval harness. Logging is already running in Langfuse. You build the layer that makes it actionable: golden datasets of representative inputs and expected outputs, automated pipelines that run evals when prompts change, and metrics that surface quality problems before a customer does.
You handle data governance for AI. What customer data can go to external LLM APIs? Where does PII live, and how do you ensure it doesn’t leave the system? These questions have real contractual and regulatory answers. The engineering implementation is yours.
You build toward RAG and MCP in parallel. Both are strategic directions, not fixed sequencing. As features require retrieval over larger knowledge bases, you build the pipelines. As the product strategy calls for making Softrip’s capabilities agent-callable via Model Context Protocol, you build toward that. Which comes first depends on what the product demands - both are real possibilities within the first 6 months.
You work across both products. Enterprise and SMB customers have different architectures and different tolerances for AI behavior. AI infrastructure decisions touch both, and you carry them across both products.
Critical Skills
• Strong backend engineering fundamentals. Python and/or TypeScript, API design, async systems, production databases. This is non-negotiable - the foundation that everything else is built on. You’ve shipped software to real users and dealt with what happens when it breaks.
• LLM API integration in a real product. You’ve built with OpenAI, Anthropic, or similar APIs beyond a demo - function calling, structured outputs, streaming, cost and latency tradeoffs in a production context.
• Agentic systems, beyond hello-world. You can reason through how agents manage state, invoke tools, and recover from failures. You’ve worked with LangGraph, LangChain, LlamaIndex, or equivalent in something that shipped or was meaningfully tested.
• Eval-first instinct. You ask “how will I know if this works?” before you start building, not after. You’ve built evaluation infrastructure, not just shipped and hoped.
• Product-minded. You care whether the feature works for the customer, not just whether it compiles. You can work directly with the AI Product Manager on what to build and why - not just how.
• Small-team operating experience. You’ve shipped in a team of fewer than 5 people and owned the full path from development to production. Former founders and consistent side-project builders - people who build because they want to, not because a job required it - are strongly preferred.
• Experience range: we’re hiring for capability and trajectory, not tenure. The right candidate might have 2 years of experience with meaningful AI systems shipped to real users, or 6 years of strong traditional engineering with serious, self-directed AI building on top - show us what you’ve shipped. Demonstrated independent experimentation - open source contributions, side projects, self-directed building - is legitimate signal and weighted accordingly.
Bonus Skills
• Hands-on RAG experience: built a working retrieval pipeline, made real decisions about chunking strategy and vector store tradeoffs
• MCP server experience, or direct working knowledge of Model Context Protocol architecture
• Langfuse, LangSmith, or Arize Phoenix experience beyond basic setup
• Comfort debugging and improving system prompts - enough to not need help with it
• Cloud deployment experience (AWS, GCP, Azure), Docker, basic CI/CD
What Success Looks Like
First 30 days: At least one AI feature is live with beta customers and being evaluated - not just shipped. We have extensive documentation and direct customer access to accelerate your ramp. The eval harness is running. You understand both products’ data models well enough to make sound decisions about context assembly, and you’ve identified the next engineering priorities.
6 months: Multiple AI features are live, measured, and iterated. You and the AI Product Manager have a tight working rhythm. RAG pipelines and/or MCP surfaces are in development or production depending on where the product has led. Data governance is solved and documented. The architectural decisions you’ve made are ones you can defend.
12 months: The AI layer across both products is a compounding asset - each feature building on shared infrastructure and shared learnings, not starting from zero. Softrip is meaningfully more capable as an AI platform than it was when you arrived, and the trajectory is clear.
