Software Engineer, Applied AI

About Clay

Our mission is to help organizations turn any growth idea into reality.

We see growth as a creative practice, not a formula. Finding and reaching your best-fit customers takes unique ideas and constant iteration. As AI makes execution faster and tactics easier to copy, creativity is the only lasting advantage. We're already helping thousands of customers — including Anthropic, Notion, Google, and Ramp — go to market with unique data, signals, and AI research.

In 2025, we raised a $100M Series C backed by world-class investors including Sequoia, CapitalG, and First Round — and crossed $100M in revenue.

In 2026, we announced our second employee tender offer in 9 months at a new $5B valuation. We also launched a community equity round, for our customers, agency partners, and club members.

Some things to know about us:

• Our community includes 11,000+ customers, 150+ integration partners, 125+ agencies, 50+ Clay clubs, and 30k members on Slack.

• Our cultureis unique inside and outside of work. Our team members are also DJs, activists, writers, clowns, marathoners, skydivers, psychedelic therapists, social workers, and more.

• All employees can work for free with world-class coaches who specialize in creativity, management, and more.

• Our operating principles — including negative maintenance and non-attached action — guide our work. Read more about them here.

• Read about us in the NYT, Forbes, First Round Review, and more.

Hear from our employees directly on our Glassdoor page!

About the Team

Clay's product is increasingly powered by AI agents — systems that research, enrich, and take action on behalf of our users, not just generate text. Several teams are working on different layers of this: agents that execute real go-to-market workflows end-to-end, and the shared platform (harness, memory, tools, retrieval, evals) that those agents run on.

This role is a shared entry point across those teams. Depending on your background and interests, you'll be matched to a specific team as you move through the process — but every team here is working on the same underlying problem: closing the gap between an agent that looks good in a demo and one that's dependable enough to run unattended in production.

About the Role

You'll work closely with product, research-adjacent teammates, and other engineers to make sure agents aren't just capable, but reliable, steerable, and worth trusting with real work. That means the job isn't only about improving model behavior in isolation — it's about turning those improvements into measurable gains in task completion, reliability, and time saved for the people using them.

What You'll Do

Depending on the team, you might work on:

Agent products

• Design and iterate on agent behavior across real GTM workflows — for example, sourcing a Total Addressable Market (TAM) list by combining search, audience building, and enrichment into one flow

• Map manual, multi-step workflows that GTM teams do today and turn them into agent-driven flows that are as good as, or better than, a human doing it by hand

• Build and run evals that measure whether an agent actually completed the task correctly — not just whether the output looked plausible — and use them to catch regressions and failure modes

• Analyze real failures in production and systematically improve robustness, not just patch the specific case in front of you

• Work with product to take agent flows from early prototype through closed beta and into general availability, and help define what "good" looks like for each one

Agent platform & infrastructure

• Build the core agent harness that other teams build on top of, including memory systems, tool infrastructure, and retrieval architecture

• Improve agent performance through prompting strategies, tool-use design, and context construction — the layer between "the model can do this" and "the product does this reliably"

• Design guardrails and safety checks so agents behave predictably in production

• Build a cross-surface evals framework so every team building on the platform can measure quality, regressions, and performance the same way

• Build feedback loops that turn real usage and production logs into better prompts, tools, and eval coverage over time

• Support teams building their own forks or variants of the managed agent for their specific use case

What You'll Bring

• Experience building or shipping production systems with LLMs or agents — not just prototyping. This might look like prompting and tool-use design, agent orchestration, retrieval, structured extraction, or fine-tuning

• Strong backend fundamentals — APIs, databases, distributed systems — since agent features still need to run reliably inside real production infrastructure

• Experience with model or agent evaluation: designing evals, measuring regressions, or turning fuzzy quality questions into measurable signals

• A systems-and-outcomes mindset — you care about whether the product actually works for users, not just about model metrics in isolation

• Comfort debugging messy, real-world failures and a bias toward shipping and iterating quickly in a space where best practices are still being figured out

Nice to Haves

• Experience with agent frameworks, tool-calling systems, or retrieval architectures (vector search, hybrid search, RAG)

• Experience building or maintaining eval/benchmark infrastructure for LLM-based systems, or running fine-tuning in production

• Experience with GTM, sales, or marketing workflows (e.g. lead sourcing, enrichment, audience building)

• Familiarity with Clay's stack: React, TypeScript, Python, AWS (Aurora/Postgres, ECS/Fargate, Lambda, OpenSearch, Elasticache/Redis), Terraform, Datadog

• A growth mindset — we're building a team that's curious, open-minded, and happy to invest in each other's learning, not just their own

Here's the section-by-section breakdown of what changed:

Intro → split into "About the Team" / "About the Role"

• Same content, but restructured into two labeled sections (borrowed from OpenAI's posting format). "About the Team" keeps the original framing but adds a concrete summary of what the platform layer covers (harness, memory, tools, retrieval, evals) and reframes the shared goal as "closing the gap between an agent that looks good in a demo and one that's dependable enough to run unattended in production" — this line is doing the real work of the OpenAI-style framing.

• New "About the Role" paragraph added entirely — wasn't in v1. Sets up that the job is about turning model improvements into measurable product outcomes, not just model quality in isolation.

What You'll Do → Sculptor bullets

• Added two new bullets: "Analyze real failures in production and systematically improve robustness" and folded "define what good looks like" into the last bullet. Both are new, pulled from the OpenAI posting's emphasis on production failure analysis rather than just building/shipping.

• Softened "Design and ship agents" → "Design and iterate on agent behavior" (iteration language, not just one-shot shipping).

What You'll Do → Platform bullets

• Added "Improve agent performance through prompting strategies, tool-use design, and context construction" — new bullet, directly from the OpenAI posting's language about the model-to-product gap.

• Added "Build feedback loops that turn real usage and production logs into better prompts, tools, and eval coverage" — new bullet, same source.

What You'll Bring

• Added "structured extraction" as an example technique (from Ramp's posting).

• Reworded the evals bullet to be broader ("designing evals, measuring regressions, or turning fuzzy quality questions into measurable signals") rather than just "evaluation frameworks."

• New bullet replacing the old "comfort with ambiguity" one: "A systems-and-outcomes mindset — you care about whether the product actually works for users, not just about model metrics in isolation" — this is a direct echo of OpenAI's "think in terms of systems and user outcomes, not just model metrics" line.

Nice to Haves

• Added "or running fine-tuning in production" to the evals bullet.

• Replaced the closing "Diversity of perspectives and interests" bullet with a growth-mindset line about investing in each other's learning — adapted from Figma's posting.