Brief us on a build
The interesting datasets rarely exist yet. One email starts it — here is exactly what to send, and exactly what comes back.
- Modality video / image / audio / multimodal
- Domain & content e.g. first-person cooking; on-camera speech; regional dance
- Languages / regions if it matters for your model
- Volume hours, clips or records — a range is fine
- Annotation depth raw-curated / auto-annotated / QA-verified
- License constraints e.g. permissively-licensed only
- Deadline what you are planning against
Send what you have — a half-formed spec is enough to start the conversation.
- A real annotated slice of your spec, in delivery format (WebDataset shards + JSON records)
- A mini-datasheet for the slice: composition, known limitations, provenance
- QA numbers measured on the slice itself — not quoted from a brochure
- A loader snippet that runs against it as delivered
then Pilot batch in 1–2 weeks: staged delivery with an accuracy SLA verified on an agreed golden set. After the pilot — full build or a recurring feed with the same schema and per-delivery QA reports.
Agreed up front, in the contract
- Private by default. Mutual NDA, no public case studies, no client names or logos anywhere.
- Exclusivity is explicit. A domain lock, or a fully bespoke corpus that is never resold — written into the contract, not decided later.
- Acceptance is measurable. Delivery is staged against golden-set metrics agreed before the pilot — you verify quality before the build completes.
What we can build for you
Shapes we scope against, not a client list — engagements are private and never published.
First-person manipulation clips curated to your domain mix and annotated end to end. For scale: discovery sustains 126 videos/hour at a 90% fetch success rate, and the full annotation stack costs about 134 GPU-minutes per hour of delivered content — so volume is a scheduling question, answered against your deadline in the brief.
A small free sample matched to your spec — enough to run through your pipeline and measure before any commitment.
The same curated corpus refreshed on a schedule: stable schema, fresh data, QA report with every delivery.
Images from explicitly CC-licensed sources only, each record carrying its exact license code, attribution and fetch time, with a dedup audit against the public corpora — for teams whose rights story has to survive due diligence.
Natural questions over real footage where every answer is anchored to the time interval that supports it, leakage-screened against public training corpora so it can be used for evals.
Audio, documents, multilingual speech, a domain nobody has bothered to collect — the funnel, annotation stack and QA are the same whatever the modality is.