Skip to content
Ossian
Try it

For research teams whose findings get challenged

Analyse your data with unparalleled confidence.

No hallucinations. No BS. A transparent workflow that traces every key decision back to its raw data source.

Open the live prototypeSee it working
  • 16,555 real units cleaned
  • κ = .81 dual-coder agreement
  • 0 silent drops
  • Steps 2–3 live today
Interview 07

“I saw twelve panels and closed the tab.”

Survey Q11

Transparency: lowest item, 2.8 / 7

G2 review

“Powerful, but the first hour is a wall.”

A/B study

Guided 63% vs control 41% activation

One defensible finding

“Onboarding overwhelm drives mid-market churn.”

Traceable to all 4 sources

The product, not a promo

Watch a cleaning decision get approved.

ossian.onrender.comReview · Step 3
unit #023 · row 24
strip_htmlremove_emojitrim_whitespace
Original — never overwritten
␣␣<b>Great product!</b> &amp; fast shipping 🔥😂␣␣
Cleaned — deterministic, reversible
Great product! & fast shipping
ApproveReject
cleaning_action #482 · before/after stored · reversible

Recreation of the live Review screen · demo data

Try it yourself

The ten seconds that matter

Click the decision. See everything behind it.

Ship guided onboarding

Go

activation +22 ptsHigh confidence68% prevalence

MethodReflexive Thematic Analysis · 4 sources · confirmed by A/B
ReliabilityCohen's κ = .81 · threshold 0.70 passed
FrameworkDeclared before analysis · technology-adoption (v2)
Evidence — click to open
Source
interviews/p07.vtt · unit #1284
Coder A1
onboarding_overwhelm
Coder A2
onboarding_overwhelm
Source
experiments/ab_onboarding.csv
Test
Two-proportion · h ≈ 0.45
Result
p < .001 · pre-registered
Source
reviews/g2_export.csv · unit #0442
Coder A1
first_run_friction
Coder A2
onboarding_overwhelm

Sample project · demo data

Other tools hand you a summary. Ossian hands you the receipts.

Before / after

16,555 units in. 10,039 you can defend.

17 real sources, end to end. Every removal logged and reversible. Nothing silently dropped.

16,555raw units, one pass
  • 10,039Clean, defensible
  • 5,611Duplicates removed
  • 905Empty removed
  • 74Flagged, kept for review

Rules, applied

Mojibake repaired

repair_mojibake
don’t buy this, café is “terribleâ€
don’t buy this, café is “terrible”

HTML and emoji stripped

strip_htmlremove_emoji
<b>Great product!</b> &amp; fast shipping 🔥😂
Great product! & fast shipping

Speaker and timestamp to metadata

separate_timestampseparate_speaker_label
[00:04:12] Interviewer: How did that make you feel?
content: “How did that make you feel?” · speaker: Interviewer · time: 00:04:12

Dates standardized to ISO 8601

standardize_datestandardize_date_epoch
07/11/2024 | 11th July 2024 | 1720694400
2024-07-11 | 2024-07-11 | 2024-07-11

7 formats · 15+ rules · 100% audit-logged · 0 silent drops

The pipeline

Your method, end to end.

The real architecture. Stage one runs today; the rest ships weekly. Withheld labels travel in our YC application, not on this page.

  1. User
    imports the data
    Data cleaning process
    deterministic rules
    Clean data
    approved units
    Data chunks
    segmented for coding
    Data storage
    pointer to raw · pointer to clean

  2. Reflexivity log
    assumptions on record
    Framework and RQ provided by user?
    Yescarry the user's framework forward
    Nofall back to default framework
    Default framework
    proposed, not imposed
    Data storage
    framework version pinned

  3. Agent 1 labeling / code
    Agent 2 labeling / code
    Independence and comparability check

    Structure shown, labels withheld — full architecture under NDA.

  4. Cohen's κ ≥ 0.70?
    Yes
    No
    Data storage
    Every stored chunk carries

    Structure shown, labels withheld — full architecture under NDA.

  5. LLM explanation layer
    Chunk fetcher
    Source aggregation
    Claim normalization
    Claim generation
    Theme grouping
    Evidence builder

    Structure shown, labels withheld — full architecture under NDA.

  6. Adaptive triangulation decision
    Fetch approved chunks with label + business context
    AI policy generator
    Policy validation
    Business context

    Structure shown, labels withheld — full architecture under NDA.

Stages 2–6 · shipping weekly

Bring your own framework.

RTA, a technology-adoption model, your team's own scheme — Ossian runs the method you already defend in review.

A nudge, not a handoff.

The tool asks before it decides. You keep interpretive control; you lose the busywork.

Defensible by construction.

Every change is deterministic, logged, reversible. The original is never overwritten.

Method

Rules clean. AI suggests. You approve.

01In build · step 4

Your lens is declared before the analysis.

A reflexivity log opens every project — your framework, your question, your assumptions — and that version rides into every claim.

02In build · step 5

Two independent coders have to agree.

Two model families code your data twice. Below κ = 0.70, the analysis stops and asks you.

03Live

Every claim carries its receipts.

Source ID, both labels, agreement score, framework version. Provenance is the data structure, not a citation.

04Live

Your data can stay inside your walls.

Run Ossian against a local model. Transcripts and PII never leave your infrastructure.

Team

Built by the people who needed it.

Md Rashedul Islam
Salehin Seyam
Muhammad Minhaj
SN
Fattahia Shafihun Nafim
Abdullah Al Mazid
Md Rashedul Islam

Founder

Muhammad Minhaj

Chief Engineer · Ex-Google

Fattahia Shafihun Nafim

Product Designer

Salehin Seyam

Assistant Engineer

SUM Neelim

Assistant Engineer

Abdullah Al Mazid

Assistant Engineer

Analysts aren't afraid of more work. They're afraid of losing control. Ossian gives the control back.
Open the live prototype

Working prototype — Steps 2 and 3. Import, clean, approve.