SlipstreamJobs tracks this role from the company's public career site. Apply directly on the employer's site.
Tessera Labs is an enterprise AI platform that helps large companies transform their operations—processes, data, and code—in weeks rather than years. The company raised $60M Series A from Andreessen Horowitz and others.
As Research Scientist, you will set and pursue a research agenda on reliable long-horizon agentic behavior in real enterprise environments. This is a rare opportunity: Tessera has actual customer landscapes, execution traces, and verifiable outcomes that most AI labs lack.
Key responsibilities:
- Define your own research agenda on agent reliability in complex, undocumented enterprise systems. You decide which questions matter and defend the choice.
- Invent and validate post-training methods for agents performing multi-step transformations: reward design where verification is partial or delayed, RL formulations for long-horizon planning and tool use, curriculum and data strategy.
- Design memory architectures for agents operating across 40+ steps and multi-day runs. Solve how agents persist, structure, retrieve, and revise their understanding when the world changes—and how to train them to use memory effectively rather than ignore it.
- Own evaluation methodology: develop metrics that predict real customer-observed correctness and expose where cheap automated proxies fail.
- Study multi-agent failure modes—error compounding, planning under partial observability, delegation and verification—and design systems to prevent them.
- Work on the verification problem: how agents establish that a change preserved behavior when no test covers it.
- Develop enterprise representation: turn process, data, and code into ontologies or knowledge graphs agents can reason over reliably as underlying systems evolve.
- Investigate post-training compute and data trade-offs across affordable model scales.
- Collaborate with Research Engineers to scale findings and with product to ship results.
- Publish papers, technical reports, and open-source artifacts; represent Tessera's research externally.
- Mentor engineers and raise the research bar through rigorous experiment design review.
This role emphasizes inventing methods rather than applying existing ones. Work is grounded in real verification signals (does the transformation build, pass tests, behave equivalently?) rather than preference models. The research setting is unusually well-suited to studying long-horizon agentic behavior at scale. Tessera post-trains open-weight models on rented compute; if your research requires 10,000+ GPUs, this is not the right fit.
