About Human Archive

Human Archive is a research lab backed by Y Combinator focused on modeling human embodied intelligence.

Humans are the most sophisticated biological systems we have ever observed, yet we still do not fully understand ourselves. Research into human physical intelligence — including the human hand, proprioception, and vision — remains largely unsolved. Our mission is to recover human embodied intelligence as a learned model. To achieve this, we build custom hardware products, deploy them globally at scale, and publish research. Today, our data is used for robotics and world modeling, but the broader opportunity is advancing scientific research into intelligence itself.

Founded by Stanford and UC Berkeley researchers, we are lean, deeply technical, and operate at extreme speed, taking on unglamorous and conventionally impossible problems that directly unlock step-function gains in model capability.

The deployment of capable humanoids at scale will permanently redefine human labor. Undesirable physical work will disappear, and human effort will shift toward a new era of abundant creativity.

We are building the infrastructure to accelerate that transition by assembling the Human Archive mafia. You will own meaningful systems from day one and see your work directly impact model capabilities. This is a once-in-a-generation inflection point. If you want to help reshape physical labor and work on problems that matter at civilizational scale, join us.

About the Role

We build and run the software platform behind a large-scale data collection and annotation operation: the system of record that tracks what was collected and by whom, the tooling annotators work in, the ingest integrations that move data from physical media into cloud storage, and the reporting that tells us whether any of it is working.

These are internal systems with real users — operators in the field, annotators working through millions of items, ingest stations moving terabytes, and machine clients writing back automatically. They are small enough that you will own whole systems and serious enough that correctness matters: a duplicated record or a dropped identifier means data that can no longer be traced to what produced it.

We want a strong generalist. You will design schemas, build APIs, write the interfaces on top of them, and deploy and debug the whole thing in production. Nobody here hands you a spec.

What you’ll do

• Design and build services end to end — schema, API, and the internal interfaces on top — primarily Python, FastAPI, PostgreSQL and React

• Build a bunch of dashboards and internal web apps to track metrics, KPIs and deployments

• Own data models where correctness is load-bearing: explicit table grain, immutable external identifiers, idempotent writes, derived rather than duplicated state

• Build tooling for human review workflows, including putting model output in front of people to correct rather than having them start from scratch

• Build and maintain integration surfaces for machine clients — automated processes that read and write over authenticated service endpoints

• Deploy and operate on AWS (object storage, CDN, DNS, identity, load balancing), and lead debugging when something breaks across those layers

• Do the performance work: query profiling, indexing, caching, and the aggregate views that keep reporting responsive

• Work directly with the operations team — the people using what you build are down the hall, and their workflow constraints are the requirements

What we’re looking for

• Strong Python in production, with FastAPI or a comparable framework, and real PostgreSQL depth: schema design, indexing, query performance, migrations

• Enough front-end capability to own the client as well as the server — React and TypeScript. You need not be a design specialist, but you should ship a usable internal tool without waiting for one

• API design judgment: versioning, idempotency, batch endpoints, and failure semantics that return per-item results rather than failing everything

• Data-integrity instincts. You should be able to explain why deriving a value beats storing a second copy of it, and why a client-generated identifier solves a retry problem

• Ability to debug across boundaries — application, database, network, cloud service — and to isolate a fault to a layer before proposing a fix

• Fluent use of AI tools — Claude, Claude Code, Cowork or equivalent — as a normal part of how you build. We care that you ship well with them and that you review what they produce: you own the code that lands, whoever or whatever drafted it

• Comfort with ambiguity and ownership: choosing an approach, writing down why, and living with the consequences

• Clear technical writing

Nice to have

• Serving or integrating model inference in a production path — batching, latency, versioning, handling low-confidence output

• Annotation, labelling, or any human-in-the-loop review system

• Clients that must tolerate unreliable connectivity and reconcile cleanly afterwards

• Media or sensor data at scale — video pipelines, large binary object storage, structured time-series recording formats

• Infrastructure as code, CI/CD, or observability — we are light here and would welcome someone who raises the floor

• Tooling that exposes internal data to language models