HUD

Commercial

high confidence

www.hud.ai ↗ · Status: active confirmed · Founded 2025

HUD (Human Union Data, YC W25, formerly hud.so) calls itself the platform for building high-quality post-training datasets. It offers an MIT-licensed SDK for defining RL environments, tasks and graders across coding, browser, computer-use and robotics; a hosted training and eval platform; and DataVendor, a marketplace where independent builders sell RL environments and data to AI labs. It announced a $16M Series A led by Standard Capital in June 2026.

Key facts

Headquarters

San Francisco, USAconfirmed cite

Headcount band

11-50reported cite

Total raised

$16Mconfirmed cite

Last round

Series A, $16M (announced 2026-06-18), led by Standard Capitalconfirmed cite

SOC 2

claimedreported cite

What they sell

environments + platform (SDK, training and eval platform, DataVendor environment and data marketplace)confirmed cite

Open source

yesconfirmed cite

Deployment

managed-hosted + self-hosted (open-source SDK run locally or on owned infrastructure; hosted platform with REST API, TrainingClient and model gateway; Docker, Modal and Daytona runtimes; Fireworks RL integration documented)confirmed cite

What's new

8 sourced updates · last 6 months

5 Oct 2026hire Hiring surge: about 30 open roles across SF, Singapore and remote (Asia and North America)
Snapshot taken 2026-10-05. The YC jobs page lists 31 roles and hud.ai/careers lists 30. They include robotics evals and data research engineers, a Forward Deployed Research Engineer, Product Manager (Marketplace), Legal Operations Lead, and a Senior Security Engineer who will 'Own SOC 2 and customer trust'. Offices are in San Francisco and Singapore.
Why it matters: Shows a fast post-Series A scale-up, with a security and compliance build-out and new robotics capacity under way.
25 Sep 2026product SDK pre-release adds managed Claude Code and Codex CLI agents
hud-python v0.6.19.dev1 (2026-09-25) introduces managed Claude Code and Codex CLI agents as AgentTypes. This is a pre-release; the latest stable version is v0.6.18 (2026-09-20).
Why it matters: Teams can evaluate or collect rollouts from off-the-shelf coding-agent CLIs inside HUD environments.
22 Sep 2026partnership HUD and Browserbase publish joint research on browser-eval reliability and open-source four LiveWeb sample tasks
Browserbase's post (2026-09-22, by Parth Patel, Ilya Kuhta and Shriya Lolabattu) and HUD's research post (dated 2026-09-28) describe building frontier RL browser tasks on live sites with Browserbase Verified mode (Web Bot Auth, recognised by Cloudflare and reCAPTCHA). HUD open-sourced 4 tasks from its LiveWeb set (Ladybird, Chairs, AirNow, Costco login) under MIT. Results varied widely by task and model.
Why it matters: Deals with live-web eval drift, grader fairness and agent identity, which matters to teams building browser-agent evals on real sites. Both companies confirm the collaboration.
22 Sep 2026research HUD publishes 'What is a good task?' guidance on task realism, fairness and difficulty
Ilya Kuhta and Reinis Salaks set out a framework for RL and eval tasks: they should be realistic (real prompts and environments, no leaked rubrics), fair (graders that avoid false positives and negatives) and hard because of the problem itself, not because of vague prompts or artificial constraints.
Why it matters: Shows the quality bar HUD says it applies to tasks, which helps anyone buying or reviewing environments from its marketplace.
Aug 2026product SDK adds Docker Compose environments, Harbor interoperability, a Fireworks RL cookbook and a unified robot stack; Daytona RL cookbook follows in September
hud-python v0.6.13 (2026-08-07) added Compose as a runtime option, Harbor integration and the unified robot stack. v0.6.14 (2026-08-19) added a Fireworks Serverless RL cookbook that trains a Fireworks LoRA adapter on HUD rollouts. A Daytona RL cookbook ('Run 256 parallel graded rollouts on Daytona and train on the results') landed in the v0.6.16 pre-release cycle (v0.6.16 shipped 2026-09-10). The PyPI package was renamed to `hud` in v0.6.10 (2026-07-15). These are documented integrations; no formal Daytona or Fireworks partnership has been announced.
Why it matters: Gives teams a documented path from environment to parallel rollouts to RL training on third-party sandbox and training infrastructure, and to Harbor-format task portability.
31 Jul 2026benchmark HUD releases Assemble Bench, an open robotics benchmark for contact-rich assembly, plus the CG-DAgger method
Authored by Lukass Kellijs and the HUD team. It has 14 simulated tasks in 4 suites (round pegs, square pegs, gears, nut threading), modelled on the NIST Assembly Task Board and built on Isaac Lab Arena. It evaluates the pi0.5 DROID checkpoint and introduces Code-Gated DAgger, where code detects failures and a scripted expert recovers in place of a human. HUD released an MIT repo, a 1,355-demo dataset on Hugging Face and two pi0.5 checkpoints.
Why it matters: Shows HUD moving into robotics and VLA evaluation and training, beyond computer-use and coding agents.
18 Jun 2026funding HUD announces a $16M Series A led by Standard Capital
HUD announced on X on 2026-06-18 that it had raised $16M. Its blog post, dated 2026-06-30, names Standard Capital as lead, with Y Combinator, Exceptional Capital, Liquid 2 Ventures and Twenty Two Ventures, plus angels Dylan Patel (SemiAnalysis), Roon (OpenAI), Ivan Burazin (Daytona) and Theo Browne. HUD says 50+ businesses use the platform and that its vendor marketplace moves 'millions' a month. Both figures are self-reported.
Why it matters: This is HUD's first disclosed round amount. It presents HUD as a platform and marketplace for RL environments rather than a single-vendor data shop.
Jun 2026product HUD repositions as a post-training data platform: SDK, Training + Eval Platform and the DataVendor marketplace
The homepage and the Series A post now present three products: HUD SDK, Training + Eval Platform, and DataVendor, which connects data suppliers with labs. The homepage self-reports 1.3M+ task runs, 2,500+ environments created and 15M+ inference calls, and publishes plan tiers (free SDK + Platform, Cloud, Enterprise, Academic).
Why it matters: Program managers can now use HUD both as tooling and as a channel for sourcing environments from third-party builders.

Founding team

Jay Ram

Co-founder & CEO

YC bio: 'prev: consumer apps, ml research, quant research'.

Profile ↗

Lorenss Martinsons

Co-founder & Co-CEO (previously listed as CPO)

Cognitive Science background at Yale (per FounderTrace). Wrote HUD's 2025 'HUD Autonomy' post on agent evaluation.

Profile ↗

Parth Patel

Co-founder & CTO

Focuses on evals and RL environments (per FounderTrace). Co-authored the 2026-09 Browserbase x HUD browser-task post.

Profile ↗

What practitioners should know

For researchers

Open research posts: Assemble Bench and CG-DAgger for VLA robotics (2026-07-31), LiveWeb browser-eval reliability with Browserbase (Browserbase 2026-09-22, HUD 2026-09-28), and 'What is a good task?' on task and grader design (2026-09-22).
Open releases under hud-evals on Hugging Face: AssembleBench dataset, two pi0.5 AssembleBench checkpoints, SheetBench-50, OSWorld-Verified, SpreadSheetBench, Online-Mind2Web-Tiny. On GitHub: assemble-bench, liveweb-samples, hud-ops-diagnostics (all MIT where checked).
The SDK's TrainingClient turns graded eval runs straight into training data (every rollout returns a trace_id and a reward), so you can bring your own GRPO or PPO. Documented RL paths exist for Daytona sandboxes and Fireworks Serverless RL.

For program managers

$16M Series A announced 2026-06-18, led by Standard Capital. Angels include Roon (OpenAI), Dylan Patel (SemiAnalysis) and Ivan Burazin (Daytona). Valuation is undisclosed.
Self-reported scale: 50+ businesses on the platform, 1.3M+ task runs, 2,500+ environments created, 15M+ inference calls, and a marketplace moving 'millions' a month. These are unaudited company claims.
Engagement is now a platform plus a marketplace (DataVendor), where third-party builders sell environments to labs. Check the provenance and QA of any environment sourced through the marketplace.
Published plan tiers on the homepage: SDK + Platform access free, Cloud at $0.10 per environment-hour, Enterprise custom, Academic $100 in free credits for .edu emails.
Team of 24 on the YC profile with about 30 open roles. Offices are in SF and Singapore. Lorenss Martinsons is now listed as Co-CEO next to CEO Jay Ram.
No frontier-lab customer is publicly confirmed by name. Public lab ties are the OSWorld-Verified feedback co-acknowledgement and an OpenAI employee (Roon) investing as an angel. Current homepage logos show only Sharpe and UiPath (UI-CUBE), both self-claimed.

For data ops

The SDK connects over MCP, CDP browser control, RFB/VNC computer use, SSH shell and files, and robot control over WebSocket (beta). Environments can be declared as Docker Compose projects (since v0.6.13, 2026-08-07).
Harbor interoperability is documented: Harbor tasks can be imported as HUD tasksets and Compose artifacts, and HUD tasks exported to Harbor folders.
Robotics data is published in LeRobot format. The HF org also shows a 'datavendor-test-lerobot' dataset.
PyPI package renamed from hud-python to `hud` in v0.6.10 (2026-07-15). Update pinned dependencies.
Security: the homepage claims 'SOC 2 compliant infrastructure' for the Enterprise plan, but there is no public SOC 2 report or trust center. A current YC posting for a Senior Security Engineer says the hire will 'Own SOC 2 and customer trust, including ... audits'. Ask for the attestation letter or report. The only legal document found is the DPA at hud.ai/dpa.
REST API reference and an LLM-readable docs index at docs.hud.ai/llms.txt.

Benchmarks & research

5 published

LiveWeb samples (with Browserbase)dataset · 22 Sep 2026

Four MIT-licensed live-site browser tasks open-sourced from HUD's LiveWeb set: ladybird-spot-num (Sketchfab), diamond-rose-seats (chair counting), epa-air-quality (AirNow) and costco-login.

Self-reported: on Ladybird, GPT-6 Astra averaged 93.3%, Claude Fable 56%, and Claude Sonnet 5 and Qwen 3.8 Max 0%. On AirNow, 20 of 21 runs got full marks. On Costco login, all four models scored 75%.

Open ↗Paper ↗

Assemble Benchbenchmark · 31 Jul 2026

Simulated robotics benchmark (Isaac Lab Arena, MIT licence) of 14 contact-rich assembly tasks in 4 families (round pegs, square pegs, gears, nuts), modelled on the NIST Assembly Task Board. Ships with a 1,355-episode demonstration dataset (hud-evals/AssembleBench on Hugging Face) and pi0.5 checkpoints (pi05-AssembleBench-12k, pi05-AssembleBench-cgdagger-r3).

Self-reported: after CG-DAgger rounds trained only on the 8mm round-peg task, pi0.5 improved dense reward on the 12mm and 16mm round pegs, with little transfer to 4mm. Sparse task success stayed limited at all sizes.

Open ↗Paper ↗

Production debugging RL environment (hud-ops-diagnostics)environment · 20 Jan 2026

Hierarchical orchestrator and subagent environment (Sentry, Supabase, Kubernetes, with Railway also integrated) built on 24 real production debugging tasks. Available on the platform as 'cross-service-diagnostics'.

Self-reported: a Sentry subagent fine-tuned from o4-mini with OpenAI RFT improved from 6.3% to 13% success, about 2x.

Open ↗Paper ↗

SheetBench-50benchmark · 1 Oct 2025

50 financial-analyst spreadsheet tasks (29 data preparation, 16 derivation/extraction, 5 modelling/forecasts), built with Sepal AI and finance professionals from PwC, Cisco, Charles Schwab and Fannie Mae.

Open ↗Paper ↗

OSWorld-Verified (HUD-hosted taskset; HUD credited as a feedback contributor)benchmark · 28 Jul 2025

XLANG Lab's refreshed desktop computer-use benchmark (369 tasks per Cua). XLANG thanks 'Human Data' (HUD's earlier name and hud.so domain) among institutions that 'provided feedback and participated in the fixes'. HUD hosts a runnable version but did not author the benchmark.

Open ↗Paper ↗

Products

HUD SDK (`hud` package, formerly hud-python) ↗

MIT-licensed Python framework for defining environments, tasks and graders. It connects over SSH, MCP, CDP (browser), RFB/VNC (computer use) and a beta robot WebSocket, with native Claude, OpenAI and Gemini harnesses. TrainingClient turns graded rollouts into RL training data. Install with pip install hud (renamed on PyPI 2026-07-15).launched 2025

Training + Eval Platform ↗

Hosted platform for running evals at scale, viewing traces, managing tasksets and environments, and RL training from graded rollouts. It has a REST API and an OpenAI-compatible model gateway.

DataVendor ↗

Marketplace connecting independent RL-environment and data builders with model labs. HUD says it moves 'millions' a month (self-reported).launched 2026

HUD docs and cookbooks ↗

Documentation including Harbor interoperability, Compose environments, and Daytona RL, Fireworks Serverless RL, robot-benchmark and ops-diagnostics cookbooks. An LLM-readable index is at docs.hud.ai/llms.txt.

Scale & velocity

Current headcount

~24 (per YC profile, accessed 2026-10-05)reported cite

Headcount growth

growing (YC team size 15 to 24 between 2026-06-07 and 2026-10-05; ~30 open roles)reported cite

Open roles

31reported cite

Other locations

Singaporeconfirmed cite

Distributed / remote

partial (SF and Singapore offices; most roles open to remote in Asia or North America)confirmed cite

Research depth

Has researchers

yesreported cite

Researcher count

unknown

Backgrounds

Jay Ram (CEO) - consumer apps, ML/quant research (YC bio), Lorenss Martinsons (Co-CEO) - Cognitive Science, Yale (FounderTrace), Parth Patel (CTO) - evals and RL environments, Named research authors: Lukass Kellijs (robotics), Ilya Kuhta and Reinis Salaks (evals/task design), Shriya Lolabattu (browser evals), Team described as including Olympiad medalists and ICLR/NeurIPS-published researchers (carried over from the 2026-06 baseline; not re-verified)reported cite

Capital

Total raised

$16M+ (Series A $16M plus an undisclosed seed)confirmed cite

Last round

Series A, $16M (announced 2026-06-18), led by Standard Capitalconfirmed cite

Investors

Standard Capital (Series A lead), Y Combinator (W25), Exceptional Capital, Liquid 2 Ventures, Twenty Two Ventures, Dylan Patel (SemiAnalysis, angel), Roon (OpenAI, angel), Ivan Burazin (Daytona, angel), Theo Browne (angel)confirmed cite

Valuation

unknown

Revenue signals

Self-reported: 50+ businesses on platform; vendor marketplace moves 'millions' monthly; 1.3M+ task runs, 15M+ inference callsreported cite

Date	Round	Amount	Led by
18 Jun 2026	Series A	$16M	Standard Capital	source ↗

Security & compliance

SOC 2

claimed (homepage: 'SOC 2 compliant infrastructure' on Enterprise plan); no public report or trust center foundreported cite

Other certifications

unknown

Security page

https://www.hud.ai/dpa (Data Processing Addendum)reported cite

Product

What they sell

environments + platform (SDK, training and eval platform, DataVendor environment and data marketplace)confirmed cite

Open source

yesconfirmed cite

License

MITconfirmed cite

Deployment model

Maturity

GAestimated cite

Notable customers

UiPath self-claimed Sharpe self-claimed ⚑OpenAI self-claimed ⚑Anthropic self-claimed cite

Buyer analysis

Best fit: Teams that want to build, evaluate and RL-train agents on reproducible environments (computer-use, browser, coding, robotics), or that want to buy or sell RL environments through a marketplace.

How we verified this

2026-10-05 refresh: I opened every primary source the draft cited and nearly all of it held up. The $16M Series A (Standard Capital lead, YC, Exceptional, Liquid 2, Twenty Two, four angels) is confirmed by HUD's blog post (dated 2026-06-30). An fxtwitter mirror confirms the X announcement went out on 2026-06-18, which is now the event date. The YC profile confirms the founders, Lorenss Martinsons's new Co-CEO title, a team size of 24 and Active status. The three YC avatar photo URLs return single-person JPEGs. The homepage confirms the three-product repositioning, the self-reported metrics, the Sharpe and UiPath case studies, the 'SOC 2 compliant infrastructure' claim and the legal name Human Union Data, Inc. I made these corrections: the Assemble Bench result overstated success gains (only dense reward improved); the LiveWeb figures were imprecise; the Browserbase collaboration first appeared 2026-09-22 on Browserbase's blog, now added as third-party corroboration; the Daytona/Fireworks item had an unsupported date and 'partnership' framing, so I re-sourced it to GitHub release notes (Compose and Harbor 2026-08-07, Fireworks 2026-08-19, Daytona around early September) and recategorised it as product; the SOC 2 job-posting source was replaced with the live YC posting; DoorDash was removed from customers, since the draft's own evidence showed it is gone from the homepage. I added material the draft missed: the PyPI rename to `hud` (2026-07-15), managed Claude Code and Codex CLI agents (2026-09-25 pre-release) and the officially published plan tiers. Web search was unavailable to me because the session's search budget was exhausted, so I checked for missed funding, acquisitions or frontier-lab deals only by direct fetches of HUD's blog, research index, X profile, Hugging Face org, GitHub releases and the Browserbase blog. None of those showed an acquisition, shutdown or named frontier-lab customer, but third-party press coverage could not be swept independently. Frontier-lab customer claims remain unproven.

Related vendors

Sources

www.hud.ai/blog/announcing-16m-series-a · 2026-10-05, Series A post dated 2026-06-30: $16M, Standard Capital lead, investors and angels, self-reported metrics, products, SF and Singapore
api.fxtwitter.com/hud_evals/status/2067646134332567770 · 2026-10-05, Mirror of HUD's X Series A post (x.com/hud_evals/status/2067646134332567770), created 2026-06-18 16:30 UTC; mentions $16M, YC, Standard Capital and 50+ companies
www.hud.ai/blog · 2026-10-05, Blog index: Series A post dated 06.30.2026
www.hud.ai/ · 2026-10-05, Homepage: three products, self-reported metrics, Sharpe and UiPath case studies, SOC 2 claim, plan tiers, legal name Human Union Data, Inc.
www.ycombinator.com/companies/hud · 2026-10-05, Founders, titles (CEO, Co-CEO, CTO), photos and links; team size 24; status Active
www.ycombinator.com/companies/hud/jobs · 2026-10-05, 31 open roles with locations
www.ycombinator.com/companies/hud/jobs/fi9ZRlo-senior-security-engineer · 2026-10-05, Senior Security Engineer posting: 'Own SOC 2 and customer trust, including ... audits'
www.hud.ai/careers · 2026-10-05, SF and Singapore offices; 30 roles
foundertrace.com/companies/hud_yc_w25/ · 2026-10-05, Founder backgrounds (Martinsons: Cognitive Science, Yale; Patel: evals and RL environments)
github.com/hud-evals/hud-python · 2026-10-05, SDK: MIT, 306 stars, connection types, harnesses, TrainingClient, pip install hud
github.com/hud-evals/hud-python/releases · 2026-10-05, v0.6.16 (2026-09-10) through v0.6.19.dev2 (2026-09-26); managed Claude Code and Codex CLI agents; Daytona runtime
github.com/hud-evals/hud-python/releases?page=2 · 2026-10-05, v0.6.10 (2026-07-15, PyPI rename to hud) through v0.6.16.dev1; Compose, Harbor and robot stack (v0.6.13, 2026-08-07); Fireworks cookbook (v0.6.14, 2026-08-19)
docs.hud.ai/llms.txt · 2026-10-05, Docs index: Harbor interoperability, Compose, Daytona, Fireworks, robot benchmark and ops-diagnostics cookbooks, REST API
docs.hud.ai/v6/cookbooks/daytona-rl.md · 2026-10-05, Daytona RL cookbook: 256 parallel graded rollouts; no partnership terms stated
www.hud.ai/research · 2026-10-05, Research index with dates and authors
www.hud.ai/research/assemble-benchmark · 2026-10-05, Assemble Bench and CG-DAgger (2026-07-31)
github.com/hud-evals/assemble-bench · 2026-10-05, Assemble Bench repo: MIT, Isaac Lab Arena, 14 tasks, checkpoints
www.hud.ai/research/can-you-trust-your-browser-eval · 2026-10-05, HUD x Browserbase LiveWeb research (dated 2026-09-28), results by model
www.browserbase.com/blog/browserbase-hud-frontier-evals · 2026-10-05, Browserbase's side of the collaboration (2026-09-22), co-authored by Parth Patel, Ilya Kuhta and Shriya Lolabattu
github.com/hud-evals/liveweb-samples · 2026-10-05, Four LiveWeb sample tasks, MIT
www.hud.ai/research/what-is-a-good-task · 2026-10-05, Task-design framework (2026-09-22)
www.hud.ai/research/debugging-rl-environment · 2026-10-05, Ops-diagnostics RL environment (2026-01-20): 24 tasks, o4-mini RFT, 6.3% to 13%
www.hud.ai/research/sheetbench-50 · 2026-10-05, SheetBench-50 (2025-10-01), Sepal AI
huggingface.co/hud-evals · 2026-10-05, 19 datasets and 2 robotics models, including AssembleBench and pi0.5 checkpoints
xlang.ai/blog/osworld-verified · 2026-10-05, OSWorld-Verified (2025-07-28) acknowledgement of 'Human Data' among feedback and fix contributors
www.cua.ai/blog/hud-agent-evals · 2026-10-05, Cua x HUD integration (2025-08-27/28); OSWorld-Verified has 369 tasks
arxiv.org/abs/2607.28367 · 2026-10-05, Checked and excluded: no HUD authors or mention
pypi.org/project/hud-python/ · 2026-06-07, PyPI package for hud-python
docs.hud.ai/ · 2026-06-07, Deployment model: cloud-hosted + local CLI, containerized isolated environments, OpenAI-compatible gateway at inference.hud.ai (Claude/GPT/Gemini/Grok); no SOC2/license stated in docs
www.hud.ai/case-studies/sheetbench-50 · 2026-06-07, SheetBench-50 developed with Sepal AI; finance professionals from PwC/Cisco/Charles Schwab/Fannie Mae involved
www.hud.ai/dpa · 2026-06-07, Data Processing Addendum page
www.workatastartup.com/companies/hud · 2026-06-07, Team described with Olympiad medalists and ICLR/NeurIPS researchers (via search snippet)
newsletter.semianalysis.com/p/rl-environments-and-rl-for-science · 2026-06-07, Technical description: HUD wraps software in dockerized container + MCP server exposing agent tools
www.linkedin.com/company/hud-evals · 2026-06-07, Correct LinkedIn handle for HUD (YC W25) RL-environments company
www.crunchbase.com/organization/hud · 2026-06-07, Crunchbase profile (HTTP 403 on fetch); funding not verified
x.com/hud_evals/status/1919262852088570225 · 2026-06-07, HUD X post: hiring research engineers, works with frontier labs to evaluate CUAs

Last updated 2026-10-05 · Every quantitative field carries a source and a confidence tag. Fields we could not source publicly are marked unknown, never estimated. See the methodology.