Job Description
Teal started six years ago, when AI was barely usable. We bet on the future anyway. Built a company. Crossed five million users. Real revenue, real product, real trust with the people who use us.
We did all of that with the tools we had then. Now we get to do it with the tools we have now.
The original idea was always to give every ambitious person the kind of career infrastructure companies have always had. In 2020, that meant trackers, templates, and tools that helped people work the system harder. In 2026, it means agents that work the system for them.
The next Teal isn’t a job-search tool. It’s the AI that runs your career.
We’re building agents that scout opportunities before they’re posted. Agents that prep you for interviews. Agents that negotiate for you. Agents that remember every project you’ve ever shipped, so the next role finds you. Your career data, your career memory, your career vault. Owned by you. Traveling with you from job to job, for the rest of your career.
The closest analogy is what top athletes have always had. Agents, managers, coaches. A team that handles everything around the craft so the athlete can stay focused on the craft. Everyone else has been left to do all of that alone. We’re closing that gap.
When ambitious people get this kind of leverage, the whole workforce levels up. More people in the right roles. More people growing into what they’re good at. AI as the equalizer that makes raw talent the deciding factor again.
You’d be joining for what’s next. The foundation is there. Five million users are there. The existing product continues to serve millions of people; the focus now is the new AI-native one.
You’d be one of three engineers on a small senior team, working directly with me (CEO), rebuilding the whole experience around agents that work for the user. You’ll ship to millions of people who’ll tell you, immediately, whether your thing worked.
Problems we’re working on
These are real and unsolved here, and pretty much everywhere else. If you read these and want to argue with me about how you’d approach them, you’re probably the person we want to talk to.
What does “good” mean when “good” is contextual? A resume bullet that’s perfect for one role is wrong for another. An interview answer that wins at one company gets cut at another. How do we eval recommendations whose quality is bound to a specific person at a specific moment, not a fixed rubric?
How do you keep an agent grounded in someone’s real career, when the data is messy? People remember projects wrong. Resumes leave things out. LinkedIn over-claims. Our agents have to reason over partial, biased, self-reported data and still produce something the user trusts.
Multi-step reasoning without the chatbot tax. Streaming a long agent run to a user without it feeling like watching paint dry. Showing the work without overwhelming. Making the wait feel like progress.
A career memory that compounds. Most products forget you the moment you log out. The vault has to get more useful over years, not noisier. What does the right schema look like for something that has to grow with someone for thirty years?
These are not toy problems. We don’t have clean answers. We want someone who treats them as the actual job.
What you’ll ship
AI features end-to-end. You own the API, the model orchestration, the streaming UI, the evals, the rollout. Nothing gets thrown over a wall.
Agentic workflows. Multi-step agents that research roles, analyze careers, write bullets, prep interviews, negotiate offers. Real agents, not chatbots with retrieval bolted on.
Streaming interfaces that feel instant, handle retries gracefully, and degrade well when a model does something weird.
The eval system this team needs. Harnesses, regression suites, latency and cost budgets. The thing that lets the next ten engineers ship without vibes.
Prompt and feature iteration infra. The system that lets us A/B test a new prompt without shipping a release.
Time with users. Customer calls, support tickets, user interviews. You’ll know who you’re building for, by name.
What you’ll build with
Edge-first AI infrastructure on Cloudflare. Workers, Durable Objects, AI Gateway, Vectorize, Queues, R2. Every user gets agents that run close to them, with state that survives across sessions, model routing we don’t have to babysit, and embeddings that don’t need a separate vector DB. This is one of the few production AI stacks built fully on the edge. It changes what’s possible on latency and cost.
Multi-model agent orchestration. Anthropic, OpenAI, and whatever’s next. Tool calling, structured outputs, streaming responses. Vercel AI SDK on the client. MCP servers for everything that talks to the world outside Teal. We treat the model layer as swappable on purpose. We want to be three weeks ahead of every model release, not three months behind.
The interface layer. React, TypeScript, shadcn/ui, Tailwind. Streaming everything. The agents have to feel alive. We obsess over time-to-first-token, retry behavior, graceful degradation when a model does something weird, and the moments of UX polish that turn a working product into one people actually love.
Career memory. The vault. Postgres, pgvector, structured and unstructured career data, semantic recall across years of someone’s work. This is the layer that compounds. Every agent we ship gets smarter because the memory underneath it is richer than what any other career product has access to.
Evals. The gap we need you to close. We have agents in production making subjective calls. Is this resume bullet good? Is this interview answer strong? Is this the right next role for this person? We’ve been vibe-checking those answers and we know vibe-checking doesn’t scale. Building the real system, harnesses, regression suites, latency and cost budgets, is one of the most important things the next senior engineer here will own. If you’ve felt the pain of shipping LLM features without evals, this is where you finally get to build the thing you wish you’d had.
Where the boring parts live. Rails, Postgres, the systems that have been earning their keep for six years and aren’t going anywhere. The boring parts stay boring on purpose. The agent layer is where the new bets go.
Stack
This is what’s in production today, not a commitment for what comes next. The new product isn’t bound by these choices. A number of architectural decisions are still open, including agent state, retrieval, model orchestration, and evals. The next senior engineer will help shape them.
Backend: Ruby on Rails, Python/FastAPI, TypeScript
Data: Postgres, pgvector, Redis
Async: Sidekiq
Frontend: React, TypeScript, Vite, shadcn/ui, Tailwind
Infra: AWS, plus room to shape what the agent layer needs next
AI: Anthropic, OpenAI, Gemini, LangChain/LangGraph, Langfuse, MCP servers, Vercel AI SDK
Tools: Claude Code, Cursor, Codex, OpenCode
What you bring
You’ve shipped real LLM features in production. Tool use, RAG, agents that don’t silently fail. Years of experience matters less than the scar tissue. We want someone with multi-step agents in muscle memory and opinions about which orchestration patterns work and which ones are demoware.
Full-stack with a spike. Strong in at least one of React/TypeScript or Rails. Comfortable in both. Willing to pick up whichever side of the stack the problem lives on without waiting for a hand-off.
Taste on evals. You’ve felt the pain of not having them. You have a point of view on what good ones look like, and you want to build the system this team is missing.
You use AI tools like a power user. Claude Code, Cursor, Codex, OpenCode. You ship 2-3x faster than engineers who don’t, and you can articulate why.
Craft. You care about UX, speed, and polish. You push back when something doesn’t feel right, not because it’s in the spec, because it matters.
Bias for closure, not just action. You can take a fuzzy problem and return a working thing. You finish what you start.
Customer proximity. You want to talk to users, read tickets, sit on sales calls. The best features come from the closest listening.
US-based, no sponsorship available.
Bonus points: vector DBs in production · MCP servers · eval frameworks (Braintrust, Langfuse, homegrown) · experience in careers / HR / recruiting products
Comp & benefits
$160k–$240k base, room for a standout
Meaningful ISOs. Early equity at a company with real revenue, not vapor
AI tools budget. Expense whatever helps you ship faster. Claude Max, Cursor, Codex, whatever’s next
Fully remote (US) · remote office stipend
Flexible PTO (20+ days suggested), 12 weeks paid parental leave, 80–100% health coverage, 4% 401(k) match
Annual team offsite
Career development stipend
1-month severance if it doesn’t work out
How we work
Weekly ship, no roadmap theater. No PMs between engineers and users. No ticket queue. No quarterly planning. You see the problem, you ship the fix, you watch the result.
AI-native by default. Claude Code, Cursor, Codex, OpenCode. We expect a 2-3x velocity multiplier from someone fluent with this. If you’re not yet, you will be inside a month.
Slope over pedigree. The fastest learners here have no interest in being the smartest person in the room. They want to be the most useful one tomorrow.
Direct access to me and to customers, always.
Some of our thoughts on careers
The Age of Agency
Contextual Intelligence: The Skill That AI Demands
Process
Apply with a GitHub repo, side project, or an AI-native thing you’ve shipped → founder chat with me → 1.5-hour max take-home coding exercise → review and discussion session → values chat with the team → paid project week where we ship something together.
Build what Teal was always going to become. Let’s talk.
