Mine production workspace data, with consent and redaction safeguards, to create benchmarks and labeled datasets.
Design label schemas, coordinate rigorous labeling, and build evaluation harnesses.
Run experiments on offline data and carry successful work into production features.
Evaluate vendors and research teams for evaluation, observability, and continual-learning tools.
Summarised from the posting by AI and can be wrong. The full description is the source.
Requirements
Experience
Strong applied research background with depth in model evaluation, benchmarking, and/or failure analysis
Work samples or code demonstrating skills such as evaluation frameworks, benchmark suites, failure-analysis reports or tooling, or labeling infrastructure
Comfort working with incomplete, ambiguous real workspace data and edge cases
Skills
Familiarity with simulation (preferred)
Familiarity with human-in-the-loop evaluation (preferred)
Familiarity with Cartridges and related context/memory-compression work (preferred)
Familiarity with memory for multi-party, long-running settings (preferred)
Familiarity with agent observability standards, such as setting up Langsmith or similar (preferred)
Your match
Sign in and add your resume to see which of this job's requirements it shows.
Saved to your account, encrypted. The AI that reads it doesn't keep it or train on it.
Benefits
Dental & vision
Dental and vision insurance
Equity
Generous equity grant vested over 4 years
Wellness & mental health
Free Equinox membership and other health perks
Also: health insurance
Full job description
From Ando’s careers site
Ando is a messaging platform where AI agents take on work alongside their human teammates. We’re rebuilding Slack from the ground up around two core ideas: durable memory and agents as first-class participants.
We have real, longitudinal, multi-party workspace data, and human / agent users whose behavior tells you whether the systems you're designing are actually meaningfully improving. Working with live business communication means permissions, redaction, and security are things we have to think about as well. If you want to have your research come into contact with reality, Ando is the place for it.
Some concrete research problems on our plate right now:
Show full descriptionShow less
Proactivity: An agent embedded in a team's channels has to decide, message by message, whether to ignore, quietly track, or intervene. Evaluating that judgment means building benchmarks where the ground truth includes silence. Existing agent benchmarks are almost entirely reactive.
Memory: What should a workspace agent remember across weeks and months of participation, in what representation, and how do you measure whether memory is helping versus hurting agent usefulness?
Continual Learning: How much does basic memory, retrieval, and context impact agent performance, vs where do we genuinely need RL and continual learning?
What you’ll be doing at Ando
You'll be one of the first members of a small research team, working close to both the data and the product.
Build evaluation from real data - Mine production workspace data (carefully, with consent and redaction pipelines you'll help design) into benchmarks and labeled datasets. Design label schemas, run labeling with real inter-rater rigor, and build the harnesses that make expert judgment cheap to capture and hard to corrupt.
Run experiments that ship - Initial work happens on offline workspace data; the destination is production systems used by every Ando customer. The distance between "the benchmark improved" and "the feature shipped" should be weeks and you'll own both ends.
Decide what we build versus who we partner with - We won't do everything in-house. Part of the job is evaluating frontier vendors and research teams (eval infrastructure, observability, continual-learning tooling) and choosing who we build with.
Publish when we have something real - We expect the team to publish as we make meaningful progress. But ideas are not Ando's moat; execution, product quality, and customer experience are. Research here is in service of customers first, and the publications will be better for it: they'll describe things that actually worked on real data.
WHAT WE'RE LOOKING FOR
Strong applied research background, with depth in model evaluation, benchmarking, and/or failure analysis. You've built evals you trusted enough to make decisions with.
Evidence over credentials. Work samples or code that demonstrate the skills: eval frameworks, benchmark suites, failure-analysis reports or tooling, labeling infrastructure. Show us something you built to find out whether a system actually worked.
Strong technical communication. You can explain complex ideas simply and hold high-bandwidth, generative technical conversations with researchers and with our product team.
Comfort with mess. Real workspace data is incomplete, ambiguous, and full of edge cases that break clean abstractions. You treat that as signal, not noise.
Bonus: familiarity with simulation (Park et al. https://arxiv.org/abs/2304.03442), human-in-the-loop evaluation (Scale HIL leaderboard https://labs.scale.com/leaderboard/hil), Cartridges https://arxiv.org/abs/2506.06266 and related context/memory-compression work, memory for multi-party long-running settings, or agent observability standards (setting up Langsmith or similar).
Hiring process
1. 30 min intro call
2. Technical conversation - walk us through an app you've shipped