Redwood City
New York, NY
Our research teams investigate search, source grounding, agents, local inference, memory, model routing, human–AI interaction, and domain operating systems. Data reliability remains a pillar — it is not the entire lab. Research ships as Ask Vector, models, and products, not slide decks.
Ground every answer in a source — warehouse metadata, documents, or official records.
We study how to retrieve the right artifact before a model speaks: dbt runs, freshness SLAs, household ledgers, and public-sector notices. Search here is not a generic web box. It is source-first retrieval so Ask Vector can cite the table, the run, or the document instead of inventing one.
View publications →Research, compare, build, and monitor as planned steps — with approval before writes.
The Agents team designs how Vector decomposes a question into lookups, comparisons, and proposed actions. Plans stay inspectable. Destructive work requires review. This is the shared workflow layer behind Ask Vector, VectorData war rooms, and VectorChat tools.
View publications →Scout on-device — private, offline after download.
We research browser and on-device runtimes so Scout can answer without leaving the machine. Local search is a first-class path, not a fallback — then Beacon and Atlas take over when the user opts into cloud routing.
View publications →Workspace, household, and conversation context that stays attached to the next action.
Memory research asks what should persist across sessions without becoming a surveillance store. We study scoped recall for VectorChat, VectorData war rooms, and Harbor households — what to keep, what to forget, and how a model cites its own prior context instead of inventing it.
View publications →Which Vector OS model answers — local, fast, archive, or flagship — without the user managing a zoo.
Auto routing is a research problem, not a dropdown. We evaluate when Scout is enough, when Beacon should guide, and when Atlas should take the long answer — one family, three tiers, one Ask Vector entry point.
View publications →Faces, voice, approval gates, and interfaces people can interrupt.
The Interaction team studies how people work with Ask Vector and VectorChat: when to speak, when to type, when a write needs approval, and how a face or command center changes trust. Multimodal interfaces — voice, live workspaces, plazas — are treated as research surfaces, not decoration.
View publications →How one intelligence layer becomes VectorData, Harbor, AmericaOS, and the rest.
A domain OS is a product interface on shared models: warehouse reliability, household finance, public-sector matching, local plazas, and sites. We research what must be product-specific and what must stay in Ask Vector so the portfolio feels like one company instead of unrelated startups.
View publications →Freshness SLOs, error budgets, and monitor design for tables — not servers.
We treat stale data as an incident class with measurable customer impact. The Reliability team studies how upstream breaks propagate through dbt graphs, how alert fatigue erodes trust, and how Fix-first ranking changes on-call behavior. Our goal is a warehouse where one root cause produces one calm signal — with blast radius, owner, and next action attached.
View publications →The cost of wrong numbers exceeds the cost of wrong queries.
The Economics team quantifies how BigQuery spend, dbt run failures, and pipeline lag compound into executive risk. We publish models for the hidden cost of stale dashboards, per-monitor pricing traps, and when metadata-only observability beats row sampling on total cost of ownership. Research here becomes pricing, audit playbooks, and ROI narratives in VectorData Enterprise.
View publications →Every alert should explain blast radius — which models, dashboards, and KPIs break downstream.
Metadata-first lineage is the grounding layer for Ask Vector. We investigate how dbt artifacts, OpenLineage events, and catalog metadata combine into explanations humans trust. Interpretability for us means no hallucinated table names: every recommendation cites the run, model, and downstream consumer that would break if you ship the fix.
View publications →dbt artifacts, INFORMATION_SCHEMA, and OpenLineage as structured signals — not generic log pipelines.
Built for platform engineers who own the warehouse. This team researches ingestion patterns that scale without copying customer rows, how to score warehouse health from job statistics alone, and how to ship reliability features behind flip switches an anchor client can turn on in a day. The output is VectorData's connectors, health scores, and the hybrid command center.
View publications →How we score warehouse health, rank Fix-first queues, and group noisy signals into one calm incident.
We publish scoring rubrics, evaluation scripts, and bake-off questions as we ship — not after a black-box launch. Open methodology covers demo scripts that surface theater vs operations, how we structure parity with observability incumbents, and what we will and will not claim before a model is production-ready.
View publications →
A bake-off script that surfaces noise, access, and price before anyone shares a service account.
Adding monitors feels productive. Ranking the break that unblocks everything else is productive.
Product leaves, solutions, learn guides, and compare pages — information architecture with a point of view.
The failure mode is not that Datadog cannot chart something. It is buying a firehose and still missing the stale mart.
VectorData is a paid product. Harbor is coming soon Fall 2026. AmericaOS is live. Labs stay labeled as labs.
| Date | Category | Title |
|---|---|---|
| August 10, 2026 | Open methodology | Questions for every data quality demo |
| August 9, 2026 | Reliability science | Fix-first beats more monitors |
| August 8, 2026 | Data platform engineering | How we structure the site like Datadog |
| August 7, 2026 | Open methodology | Datadog for warehouses is usually wrong |
| August 6, 2026 | Open methodology | Building Vectornosis in public |
| August 6, 2026 | Lineage & interpretability | Metadata-only is a feature |
| August 5, 2026 | Warehouse economics | The case for flat-price data monitoring |
| August 1, 2026 | Warehouse economics | What stale data actually costs |
| July 28, 2026 | Reliability science | Calm alerts beat noisy observability |
| July 20, 2026 | Lineage & interpretability | Metadata only: our security promise |
AI Models
Atlas, Beacon, and Scout power Ask Vector — source-grounded search across the web and your workspace. Scout is fast and lower-cost. Beacon adds reasoning and guidance. Atlas is the flagship for complex research.
We hire researchers, platform engineers, and writers who want research to ship as Ask Vector, models, and products — not slide decks.