APPLIED AI. OPEN FOUNDATIONS.
Better data for the world we work in.
Curated, openly licensed datasets for the industries that keep it moving.
01 / OUR FOUNDATION
AI becomes useful when it understands the work. We’re building the data foundations for that understanding, from forests and fisheries to customer conversations and job-specific robotics.
Our flagship is carefully curated, openly licensed training data. Outdoor apps and custom business software bring the same practical approach into people’s hands.
THE FLAGSHIP / CURATED DATASETS
Four areas of focus. One commitment: practical data that others can build on.
Training data for recognizing forest conditions, organizing field observations, and supporting the people who manage working forests.
Illustrative training applications. Explore our early banking and forestry samples below. Additional datasets are planned for launch.
Our approach starts with task relevance, careful organization, and clear documentation.
Each release identifies its source licenses, attribution requirements, and changes. Open data should come with clear terms and a traceable history.
Help established industries adopt AI and open new possibilities for practical, job-specific robotic skills.
OPEN SAMPLES / VERSION 2 · OCTOBER 2026
Two early releases built from established public sources, with readable fields, reproducible curation and room for expert review.
These samples show how we organize and document data. The underlying observations and queries are already publicly available. Our contribution here is selection, clearer structure, provenance and provisional annotations. We have not yet measured a training improvement or completed independent domain review.
Customer queries about payments, refunds, account access, identity checks, transfers and card delivery. All 77 original BANKING77 intents are included, restoring the adjacent cases that a narrow complaints-only selection misses.
| Customer query | Source intent | Readable name |
|---|---|---|
| Where is my card? | card_arrival | Card has not arrived |
| I cant prove who I am | unable_to_verify_identity | Identity verification failure |
| I want a refund | request_refund | Request a card purchase refund |
Derived from PolyAI’s BANKING77 (Casanueva et al., 2020). Original query text, intent labels and retained records’ source splits are preserved. Twelve normalized duplicate queries were removed; matching test records took priority over train records.
Plain-English intent names, 13 broad categories, a glossary, traceable record IDs and blank review worksheets. The new editorial annotations are AI-proposed and await human review. Source labels remain distinct from these additions.
Intent classification, routing experiments and label-boundary review. Training queries can seed separately labeled synthetic paraphrases. Keep the test set out of generation, fitting and prompt tuning.
CC BY 4.0 permits adaptation and commercial reuse under its attribution terms. This is not a refund-policy engine, a conversation corpus or verified banking advice. The modified test set is not the untouched benchmark. PolyAI does not endorse this adaptation.
CSV + JSONL · Intent glossary · Review sheets · Build script · Attribution, source hashes and change log
A documented sample of Maine tree observations from the USDA Forest Service Forest Inventory and Analysis database. It brings field codes, species references and measurement units together for easier inspection.
| Species reference | Recorded status | Diameter (in) | Height method |
|---|---|---|---|
| tamarack | Dead tree | 7.3 | Total estimated visually; actual measured |
| American mountain-ash | Live tree | 2.5 | Field measured |
| bigtooth aspen | Dead tree | Not recorded | Not recorded |
USDA Forest Service Forest Inventory and Analysis database, Maine extract. A deterministic selection spans inventory years, species codes and tree statuses. It favors coverage of different cases and is not a representative statewide sample.
Readable field names, decoded species and status values, units, source keys, a field dictionary and record-grounded worked examples. Source values remain traceable. Interpretive notes and examples are provisional and await domain review.
Tabular ML prototyping, field-code interpretation and grounded question-generation experiments. Related plot observations and repeated measurements need grouped evaluation splits. No ready-made train/test split is supplied.
The official FIADB catalog identifies the source database as CC0. Package notices distinguish source data from added material. This partial-plot sample cannot produce statewide estimates or harvest prescriptions. It is not a robotics action dataset. USDA and Maine Forest Service do not endorse this adaptation.
CSV + JSONL · Field dictionary · Review worksheet · Worked examples · Reproducible selection and source provenance
Each download separates source fields from proposed additions and includes space for your corrections. We welcome specific feedback from forestry researchers, field practitioners and banking service teams. Expert review will be credited only after it is completed and with the reviewer’s agreement.
Discuss a review with Hayven02 / BEYOND THE DATA
Two complementary lines of work. The same focus on software that earns its place in everyday life.
High-quality apps for hikers, climbers, and travelers. We bring machine learning and hands-on outdoor experience to tools designed around real journeys.
CRMs, websites, and custom software built around how a business actually operates. Our priority is local businesses in forestry, fishing, customer service, and applied robotics.
03 / A NAME WITH ROOTS
Krummholz is the wind-shaped growth of trees at the edge of their environment.
Our name connects the forests we study, the mountains we explore, and the way we build: adapting to real conditions. Useful intelligence needs that same grounding: data shaped by actual tasks, and software shaped by the people using it.
Krummholz is a new startup built on experience in web development, CRM systems, AI data workflows, and outdoor applications. Founder Hayven Jackson brings experience at the South Dakota State Data Center from 2013 to 2018, alongside the practical engineering work of Katsura Dataworks.
Explore Katsura Dataworks04 / MAINE ROOTS. A WESTERN HORIZON.
We’re starting in a place where working forests, coastal industries, and independent businesses make the case for practical AI tangible.
Our commitment is to help those industries gain useful new capabilities and create opportunities to build what comes next.
We are exploring an expansion of our data operations to Cheyenne beginning in late 2027, alongside our pending Wyoming LLC incorporation.
Cheyenne is central to our plans for a western base. Any expansion remains subject to funding, partnerships, and operating conditions.
THE NEXT CHAPTER
Our planned start of operations and the beginning of our venture capital fundraising.
Email me
krummholz