QUALITATIVE RESEARCH, ANALYSED IN FULL

Interviews in.
Defensible findings out.

Ossian codes your transcripts, builds the themes, and links every claim to a timestamped quote. Weeks of manual analysis, finished in an afternoon.

Request Beta Accessarrow_forwardMVP · AUGUST 2026
Private beta · Encrypted at rest · Your data never trains a model
app.ossian.app / fintech-onboarding-q3
Ossianfintech-onboarding-q3/P07_transcript.vtt4 files stagedImport and clean
Import datasetfintech-onboarding-q3✕
Drop transcripts, notes or exports here
or browse your computer · up to 500 files per batch
vttsrttxtdocxpdfcsvtsvmdjsonxlsx
OR CONNECT A SOURCEMCP
GDGoogle DriveDocs, Sheets and shared foldersconnected
ZMZoomCloud recordings and transcriptsconnected
DTDovetailProjects, highlights and notesconnect
ZDZendeskTickets, macros and satisfaction notesconnect
OTOtter.aiMeeting and interview transcriptsconnect
QUQualtricsSurvey open endsconnect
4 files staged · 1.2 MBCancelImport and clean
IMPORTCLEANFRAMEANALYSETHEMESREPORTEXPORTDrop in transcripts, notes and exports — nothing else to set up.
GROUNDED IN REFLEXIVE THEMATIC ANALYSISEVERY CLAIM TRACED TO A TIMESTAMPED QUOTEDETERMINISTIC SCORING, REPRODUCIBLE ON RE-RUN
02 · WHAT THE RUN DOES

Your sources in, an auditable analysis out.

Ossian pulls the material from wherever it already lives, works through it phase by phase, and hands back a record rather than a summary.

account_tree
Six phases, each one logged
Segmentation, open coding, clustering and evidence linking run as discrete passes, and the scoring around them runs in ordinary Python, so a re-run returns the same numbers.
format_quote
Every claim carries its quote
Nothing enters a finding without a timestamped utterance behind it, so any sentence in the report can be opened back to the source.
rule
A review queue, not a rubber stamp
Roughly five per cent of the corpus reaches you, ranked by how much your judgement changes the result. The rest is already settled.
Request Beta Accessarrow_forwardPRIVATE BETA · AUGUST 2026
hub
Connect to your favourite MCPs

Point Ossian at the tools your team already uses. Transcripts, notes and exports arrive through your own connections, and nothing has to be uploaded twice.

NotionGoogle DriveExcelSlackDropboxClaude
NotionGoogle DriveExcelSlackDropboxClaude
03 · PRIVATE BETA

Apply for beta access.

A working version of the platform arrives at the end of August 2026. We are admitting a small first cohort of researchers who will use it on real projects and tell us where it breaks.

credit_card_off
Free for the duration of the beta
Pricing is not set. You will see it before anything is charged.
forum
A direct line to the founders
You talk to the people building it, not a support queue.
lock_open
No commitment required
Leave whenever the tool stops being useful to you.
WHAT YOU ANALYSE
ROUGHLY HOW MUCH, PER MONTH
Opens your mail client, addressed to hello@ossian.app
04 · QUESTIONS

FAQ

Anything not answered here, ask us directly: hello@ossian.app.

A general model returns a plausible coding frame in a single pass and no account of its own reliability. Ossian runs the analysis as a procedure: segmentation, open coding, clustering and evidence linking are discrete, logged passes over the full corpus. The deterministic steps - confidence scoring, convergence weighting, contradiction search - execute in ordinary Python rather than inside the model, so they are reproducible on re-run. What you receive is not a summary but a structured record from which any individual claim can be traced back to the utterances that produced it.

Established, peer-reviewed procedure for human analysis, automated phase by phase rather than approximated. For interview and transcript data the default is reflexive thematic analysis across its six phases; framework mapping is available where an a priori structure is required. You supply the analytical framework and codebook you would have applied yourself - the system calibrates to your framework rather than to a generic standard.

Each high-level finding carries a score derived from measurable quantities: how many independent sources support it, whether a directed search for disconfirming evidence returned anything, and what proportion of the corpus no theme accounts for. Convergence is weighted by source independence - agreement between two sources that share a bias counts for less than agreement between genuinely independent ones. Findings below threshold are routed to review rather than reported as settled.

Verifying nothing is untrustworthy; verifying everything is impossible. The queue ranks the segments - typically around five per cent of the corpus - where your judgement most reduces the risk of a wrong finding, and asks for a decision on those alone. Each item arrives with the quote, its surrounding context, the code applied and the reason it was flagged.

Your data is encrypted at rest, and you can delete any single source or an entire project at any time. It is never used to train a model: models are prompted against your codebook, not fine-tuned on your corpus.

Pricing is not set. The beta is free for its duration, and members will see what pricing looks like before anything is charged.