knyta

knyta /ˈknŷːta/ verb, Swedish — to tie; to knot; to bind together.

Customer data arrives messy. Yours lands clean.

Knyta is an AI-assisted data-ingestion engine. It turns the CSVs, exports, and odd files other people send you into clean rows where they belong — with pipelines you can read, replay, and trust.

hello@knyta.net see how it works ↓

01 — the problem

Every file someone sends you is somebody's Tuesday evening.

Your work runs on data other people produce, and it arrives as whatever their system happened to export — headers on line four, decimal commas, three date formats in one column, a TOTAL row at the bottom. Generic ETL tools assume clean input. Reality doesn't cooperate, so someone cleans it by hand. Usually you.

orders_export_FINAL(3).csvas received 
;;Kundrapport,,
Genererad:,12/1/2026,,
,,,
ORDER NO,customer,Amount ,date
10 234,Acme AB,"1 024,50",2026-01-12
10235,acme ab,980.00,12/01/2026
10236,Beta GmbH,,13.1.2026
TOTAL,,"2 004,50",
public.ordersas delivered 
 order_no │ customer_id │ amount  │ ordered_at
──────────┼─────────────┼─────────┼───────────
 10234    │ cus_8f3a    │ 1024.50 │ 2026-01-12
 10235    │ cus_8f3a    │  980.00 │ 2026-01-12
 10236    │ cus_91bb    │ flagged │ 2026-01-13

The header noise and the TOTAL row were recognised as furniture and dropped. The space-grouped numbers and decimal commas were parsed, the three date formats reconciled. And the missing amount was flagged for review — not guessed.

02 — how it works

AI writes the pipeline once. Boring, deterministic code runs it forever.

Most "AI-powered" tools put a model in the hot path and hope. Knyta doesn't. AI is used where it's genuinely good — reading a messy file nobody has seen before and writing the code to handle it. Everything after that is pinned, replayable, and audited.

  1. A file lands on its channel

    Every sender gets their own intake channel. Files arrive however they arrive — you stop being the person who forwards attachments.

  2. Known shape? It runs the pinned pipeline no AI on this path

    Each file's structure is fingerprinted. A match runs the exact pipeline that was approved for that shape — deterministically, same file in, same rows out. A guard also compares value shapes, so a file whose columns quietly swapped meaning under unchanged headers parks for a human instead of delivering mis-mapped data.

  3. New shape? An agent writes the pipeline sandboxed

    An AI agent reads the actual file, writes readable pipeline code with an explicit validation contract, and tests it against that file through the real execution path — inside a sandbox with no credentials and no network.

  4. Nothing new touches production without approval probation

    A freshly authored pipeline starts in probation: a human reviews its first output before anything is written to a live destination. Until then it runs against a dry-run client that is structurally incapable of writing.

  5. Delivery lands transactionally

    Delivery runs as a saga: retried when provably safe, rolled back in reverse order when a step fails, and quarantined with full context when an outcome is unknowable. A file that fails validation quarantines whole — never a half-imported mystery.

03 — principles

Built the way you'd build it, if you had the month to spare.

Data ingestion fails in quiet, expensive ways. These four rules are load-bearing in everything Knyta does.

§1

Code, not a black box

Every pipeline is plain, readable code with an explicit validation contract. You can open it, diff it, and see exactly what happens to every column. No logic hidden behind an API you can't inspect.

§2

Boring on purpose

AI authors a pipeline once; after that the run path is deterministic. Re-running a file gives the same result every time. The clever part happens before approval, never during production runs.

§3

Trust is earned, not assumed

Autonomy is granted per decision type, lifted slowly after clean track records, and revoked instantly on a fault. The riskiest actions permanently keep a human in the loop. Every gate decision lands in an append-only audit log.

§4

Honest failure beats silent success

A value that doesn't validate is flagged, not guessed. A delivery whose outcome is unknown is quarantined, not reported clean. You'll never discover an import problem three weeks later, downstream, in something that mattered.

04 — who it's for

If someone else's file decides how your week goes.

We keep meeting this problem in places we didn't expect. A founder whose product runs on customer uploads. A recruiter reconciling candidate lists that arrive in a different shape from every client. An analyst whose month starts with the same spreadsheet and the same twenty minutes of cleaning.

Different jobs, identical problem: the data is someone else's to format and yours to deal with. If that's a recurring line item in your week, Knyta is probably for you — whatever your title says.

Destinations are pluggable by design: each one declares its capabilities, and the engine delivers to it with the same transactional care. Postgres is supported today; APIs, other databases, and more are in the works.

  • the file arrives as an attachment
  • the format changes without warning
  • only you know what the columns mean
  • "the export" is the integration
  • book-a-demo walls
  • per-seat enterprise pricing
  • "just send it in our template"

05 — contact

Tell us about your worst file.

Knyta is currently working hands-on with a small number of early users. If someone else's data is stealing your week, send us the ugliest file you deal with — we'll tell you honestly whether we can tie it down. We're genuinely curious where this problem turns up.