Zep manages, governs, and serves agent memory at enterprise scale. Enterprises build on Zep to run reliable, personalized agents across the business: millions of Context Graphs, served in under 200ms, inside their own VPCs and cloud deployments. Customers include Samsung, Zscaler, Twin Health, HoneyBook, and NASDAQ 100 and Fortune 500 technology companies. We also build Graphiti, our open-source context graph framework (30K+ GitHub stars).
What our customers' agents can reason about depends on the memory we retrieve and the memory we write. You own that loop. You build agents that improve retrieval. You finetune the models that extract memory, and the models that power those agents. You run the experiments and ship the result as production code.
We're hiring an engineer who builds agents and trains the models they run on.
We will measure you on whether retrieval and memory quality move, and on what reaches production.
You'll report to our founder, Daniel (2x founder, engineer, former head of ML at SparkPost), and join a team with pedigree at Scale AI, Dropbox, ActiveCampaign, DroneDeploy, and McKinsey.
How we work
We're a small, distributed team that works closely together. We pair on hard problems, review each other's designs, and treat learning as part of the job rather than something that happens after hours. We ask a lot of questions: of customers, of teammates, of our own assumptions. When we find pain, we go fix it.
We expect the same back: ask questions early, push back when you disagree, and care about the people on the other end of the API.
What you'll do
Build agents that improve retrieval: query understanding, ranking, and what to pull into context.
Finetune the models that extract, update, and consolidate memory on Zep's domain.
Finetune the models that power those agents, so the retrieval loop and the agent get better together.
Own the work from dataset creation through experiment design, evaluation, training, and a change that ships to production. Our engineers work the serving path alongside you.
Build the eval harnesses that catch regressions in retrieval, memory quality, and agent task completion before a release ships.
Write up what you find, including the results that killed an idea, so the rest of engineering can build on it.
Your first 90 days
In your first week you pick up an open retrieval or memory-extraction problem and frame it as an experiment. By day 30 an agent or a finetune you built is in the product loop, or you have ruled the approach out and written up why. By day 90 you are three cycles in and the eval harness you built runs on every release.
What we're looking for
You have shipped production ranking, retrieval, or query understanding.
You've built classical ML into a real product — logistic regression, SVMs, GBTs, single-layer perceptrons.
You've shipped finetuned models to production. You know transformer architectures and training workflows, and you work in PyTorch.
You have shipped a non-trivial agentic system to production. Not a prototype, not a thin wrapper over a chat-completion API.
You have a research methodology: dataset creation and curation, experiment design, and evaluation. You can frame an open problem and design an experiment that answers it.
You have built evaluation for retrieval, generation, or agent tasks: gold sets, offline metrics, online tests.
You write strong Python, and you have enough production AWS to ship, monitor, and iterate on what you train.
You have a Master's in Computer Science or equivalent experience.
This role is probably NOT a fit if:
Your shipped work is papers, prototypes, and demos.
Your agent work is a wrapper over a chat-completion API, with no eval and no model you trained.
You haven't finetuned a model that reached production.
You need a research agenda handed to you before you move.
Visa: US citizen/visa only.