Whether you can trace which steps produced this data.
Auto-managed metadataSteps recorded per versionRun history kept
Changes · as of v4 · 5 applied · 1 excluded
Review the fixes Ready for Raw Data proposed.
!Check 1 order with no customer infoWith no customer reference or contact, the amount alone cannot identify the customer.Not applied because it needs a manual check.ExcludedEvidence
✓Fill blanksFilled blanks by business rules so fewer rows are dropped from training.✓ Applied · v2Evidence
✓Remove duplicatesMerged rows where the same measurement was entered twice.✓ Applied · v2Evidence
✓Unify unitsUnified mixed units such as ℃ and °F.✓ Applied · v3Evidence
✓Unify time zone (UTC)Aligned measurement times to UTC so the order stays correct.✓ Applied · v3Evidence
✓De-identify personal dataReplaced inspector names and contacts so the data can be used with fewer original values exposed.✓ Applied · v4Evidence
Syntitan AgentExample
Persona survey
Survey virtual respondents similar to your customers to find answer distributions and improvements for each question.
Analysis report · PDF download
Customer success
15+ min
Download sample response CSV
Select data to analyze1/3
Demographic data is required. Survey data makes the results richer.
Customer behavior (cleaned)Ready for Raw Data
Campaign sends (cleaned)Ready for Raw Data
Demographics (organized)Ready for Raw Data
Churn training dataReady for Model
Sales history training dataReady for Model
customer_activity_2025.csvOriginal
crm_send_history.csvOriginal
institution_perception_survey_2025.xlsxOriginal
survey_2024.xlsxOriginal
Data needed for analysis
Demographic dataRequired for analysis
RequiredRespondent ageage, age_group
RequiredRespondent gendergender, sex
RecommendedPast survey responsesquestion, answer
Form types3–50 questions
Multiple choiceSingle or multiple answers
RankingPut options in order
RatingScore on a scale
MatrixRate several items on one scale
Open-endedAnswer in free text
Preview
Response distributionShows a chart and a short reading for each question type.
Standard dataUpload work files to generate test cases.
Personal or confidential dataReplace sensitive fields with stand-ins, then generate test cases.
Sample dataSee the full flow without your data.
Recent projects
ProjectTypeDatasetsStatusOwnerUpdated
Permit renewal agent validationReady for Agentpermit_rules_2026 +1WaitingJamie Park4 days ago
Inspection assistant validationReady for Agentfacility_inspectionsCompletedAlex KimOct 30
Ready for Raw DataPrepare regional orders for shared useExample
1Goal & data
26-axis diagnosis
3Review proposal
4Result & validation
What do you need to accomplish?
Tell us the purpose, source and destination. Review the proposed handling next.
Your task
Prepare reusable order data while preserving currencies and customer relationships.
Source dataUse sample data
Data type
Standard data · approved working copyPersonal or confidential data
Import method
Upload a fileCSV · XLSXUploaded datasetsDatasetsExternal sourceDatabases · cloud
data_example.csv10 rows · 8 fieldsChoose file
ⓘ DB & app connectionsⓘ File size or format issue?
Preview source
Syntitan proposesStart with the task. Syntitan proposes the steps.
Inspect the source and task conditions
Propose DTS processing and preservation constraints
Compare changes and task-specific results
You can stop at data preparation and continue with a model or agent when ready.
Product walkthrough sampleScores, transformations and evaluations illustrate the product experience; they are not customer measurements.
Ready for ModelAdapt transactions to a detection modelExample
1Goal & data
2Review related columns
3Baseline
4Result & validation
What do you need to accomplish?
Tell us the purpose, source and destination. Review the proposed handling next.
Your task
Align units and inputs, supplement scarce cases, and inspect both detection gains and false alerts.
Source dataUse sample data
Data type
Standard data · approved working copyPersonal or confidential data
Import method
Upload a fileCSV · XLSXUploaded datasetsDatasetsExternal sourceDatabases · cloud
model_example.csv12 rows · 9 fieldsChoose file
ⓘ DB & app connectionsⓘ File size or format issue?
Preview source
Task type
Binary classificationSuggested
Compares before/after performance on how well the model separates two classes.2 classesMCC · Macro F1 · ROC-AUC
Anomaly detection
Measures a detection model that finds records outside the data's normal distribution.Normal / AnomalyPR-AUC · Recall · F1
Regression
Compares before/after performance on continuous predictions.ContinuousR² · RMSE · MAE
Multiclass
Measures per-class performance of a model that predicts one of three or more classes.3+ classesMacro F1 · accuracy, etc.
Based on what you wrote, we picked binary classification as a suggestion. You can change it.
Syntitan proposesStart with the task. Syntitan proposes the steps.
Inspect the source and task conditions
Propose DTS processing and preservation constraints
Compare changes and task-specific results
You can stop at data preparation and continue with a model or agent when ready.
Product walkthrough sampleScores, transformations and evaluations illustrate the product experience; they are not customer measurements.
Ready for AgentRefund exception checkExample
1Describe the task
2Review task scope
3Cases
4Run & compare
Set up the agent's task
Add the task the agent handles and the documents it works from.
Demo: after a failed analysis
Name *
Refund exception check
Main task *
I want the support agent to handle refund exceptions correctly.
How to provide task files *
Upload fileUpload a task manual or agent MD file and we analyze the decision rules.Connect via SDKConnect the agent to read its instructions and tool settings.
Upload file *
Standard fileContains personal data
Add documents
✓ Agent MD guide uploaded. You can combine uploaded files with existing data. (Up to 10)
Upload 2Existing dataExternal sources
Drag a file here or click to uploadUp to 20 MB per file · up to 10 filesEnter text instead
Selected documents 2/10Uploaded 2
CSVrefund_manual.csv×
MDrefund-agent-guide.md×
Summary
Agent
Refund exception check
Upload file
refund_manual.csv + 1 more
Personal data protection
Off
Work StudioHomeExample
Data that was hard to feed into AI, made ready to use.Start by seeing what is wrong with your data.
From preparation to results in one placeStart without a data team.Syntitan proposes how to prepare your data first.
Examples by industryHere is what you can do with it.Pick a case close to your work and see right away how the data changes.
Prepare data for AII want to use customer order records for AI analysis, with personal details replaced.Names become stand-ins, and the same customer keeps the same value. Date and country formats are unified, and rare overseas orders are supplemented.
120 overseas order cases were supplemented. Names become stand-ins, and each customer's order flow stays intact.
Check how well AI predictsI want to know which customers will churn next month.Align fee formats and fill in missing plans, then compare how many more churning customers are found under the same conditions.
Before
CustomerPlanMonthly fee
C-01Premium55,000 KRW
C-0255000
C-03basic33000 KRW
▶
After
CustomerPlanMonthly fee
C-01Premium55,000
C-02Premium55,000
C-03Basic33,000
Check how well AI predicts
What to predict
Churn
Churning customers foundout of 200 actual
140›158
The prediction score rose from 0.710 to 0.780. You can reach out to customers likely to churn first.
Prepare data for AII want to prepare support records for an AI chatbot.Customers are linked and replaced with stand-ins even when contact formats differ. Duplicate records are merged so the same content is not learned twice.
85 duplicate support records were merged. Past inquiries are kept without being learned twice.
Check how well AI predictsI want to pick out products likely to be defective in advance.Unify temperature units and exclude sensor errors, then compare how many more defects are found under the same conditions.
Before
ProductTemperaturePressure
P-0785℃1.2 bar
P-0899991.3
P-09185°F
▶
After
ProductTemperature (℃)Pressure (bar)
P-0785.01.2
P-081.3
Sensor error excluded
P-0985.01.25 (avg)
Check how well AI predicts
What to predict
Defect
Defective products foundout of 60 actual
42›47
The prediction score rose from 0.510 to 0.580. You can inspect products likely to be defective first.
Prepare data for AII want to align scattered sales records before analyzing revenue.Date formats and currency units are unified so revenue can be combined on one basis. Missing amounts are flagged, not filled in arbitrarily.
Before
OrderDateAmount
O-112026.9.348,000 KRW
O-1209/03/26
O-132026-09-03$36
▶
After
OrderDateAmount
O-112026-09-0348,000 KRW
O-122026-09-03
Needs review
O-132026-09-0336 USD
Ready to use with AI
What was preparedThe original stays as is
Date alignmentCurrency alignment
Readiness
61 pts›84 pts
312 date and amount entries were aligned. Revenue can be combined on one basis, and missing amounts are flagged separately.
Check how well AI predictsI want to know which loans are likely to become delinquent.Unify income and amount units in KRW and fill in blanks, then compare how many more delinquent loans are found under the same conditions.
Before
LoanMonthly incomeLoan amount
L-113M30,000,000 KRW
L-12300000025M
L-1318000000
▶
After
LoanMonthly incomeLoan amount
L-113,000,00030,000,000
L-123,000,00025,000,000
L-132,800,000 (avg)18,000,000
Check how well AI predicts
What to predict
Delinquency
Delinquent loans foundout of 120 actual
84›95
The prediction score rose from 0.640 to 0.710. You can review loans likely to become delinquent first.
Handling · applied 4 · excluded 0Recommended steps are applied by default. Excluded steps are recorded as ‘Excluded’
Replace names, emails and phone numbers with stand-insThe same customer gets the same value (e.g., ‘Customer_0193’), so order history stays linked
Unify order date formatMixed formats like 2026/10/01 · 2026-10-01 into one
Supplement scarce overseas ordersOverseas orders are only 2%, so country analysis skews domestic. We add cases similar to real overseas orders and flag the added rows
Unify country labelsMixed labels like US · USA · United States into one
This stepWe’ll prepare it with 3 cleanups and 4 supplements
The original stays as is; a cleaned new version is created separately.
Syntitan prepares firstNo need to know how to read the results
Train the same AI on both versions
We compare how much better it predicts under the same conditions
We explain what changed in plain words
If it falls short, you can keep refining
Review data impactCreate projectExample
Select data
2Question
3Baseline
4Compare
What do you want to predict?
Choose what to predict, and we compare before and after under the same conditions.
This stepChoose what to predict, and we compare under the same conditions
For your goal, we compare data before and after preparation under the same conditions. If no goal fits, write your own.
Review data impactCreate projectExample
Select data
Question
3Baseline
4Compare
We predicted with the data before refining first
Results measured on your current data. We’ll refine the items below and compare with the same AI
Prediction score57%Below the threshold (75%)
Share of actual cases found52%Found 196 of 374 actual cases
Share of AI picks that were correct62%196 of 314 picks were correct
Scope
21 columns · 7,043 rows · 1,409 rows used for scoring
Column to predict: ‘Churn’
Positive side ‘Yes’ 26.5% · rest 73.5%
AI model
1 representative AI · scores for all 4 are in Expert mode
Gaps and handling to try
Handling to tryWhat we found
Drop confusing training rowsSimilar rows, different outcomes · training only
Fill blank ‘TotalCharges’ cells11 rows are blank · fill with the column median
Turn text values into numbersText values like ‘Contract’ · converted by frequency
Next stepAfter refining the data, we compare again with the same AI
Review data impactFind rows where ‘Churn’ is ‘Yes’Example
Find rows where ‘Churn’ is ‘Yes’21 columns · 7,043 rowsv2
Type
Review data impact
Question
Know who will cancel next month
Column to predict
Churn (find Yes)
Datasets
telco_churn.csv
Before/after comparison complete
Saved to recent projects. Your uploaded file stays unchanged.
ReportCompare before/after performance.
Proof KitGet before/after metrics at once
Connect a destinationConnect via MCP or API.
Predictions before/after refining
Correctly found cases rose from 196 to 288
Correctly found196▸ 288Of 374 actual cases
Newly found92Missed before refining, found after
No longer found0No cases were found before but missed after refining
What made the difference?We refined the data by dropping confusing training rows. Rows with similar values but different outcomes confuse the AI, so only the majority-side rows were removed from training. Evaluation rows used for scoring are unchanged.
Newly found row IDs1,6406,1267513,3635,9691,4733,9763,5171,9886,389+82 more
AI accuracy changeThe score rose but is still below the threshold (75%)
Recall–precision balanceA score combining the two rates on the right56.98% → 62.61%+5.63pp
Support ticket triage agent4 filesAgentv1DoneAAlex KimOct 15
Order lookup agent2 filesAgentv1DoneAAlex KimOct 1
Summary
Agent
Refund support agent
Connection
Upload files
File
3
Prepare agent testsRefund support agentExample
Select data
2Confirm use
3Check · Process
Here’s how we understood the task
Rules we found in your files. Turn off any we got wrong.
Task rules · kept 5 · excluded 0Test cases are made only from the rules you keep
Refund within 7 days of purchaserefund_policy_v3.docx
For a change of mind, the customer pays return shippingrefund_policy_v3.docx
For a damaged item, get a photo first, then refund in fullCS_support_manual_2026.md
If shipping is over 5 days late, offer a coupon firstCS_support_manual_2026.md
If rules overlap, hand off to a personFound in 12 chats in support_log_sample.csv
This stepCheck the rules the agent should follow
If your files disagree, the rule is marked for review instead of guessed.
Prepare agent testsRefund support agentExample
Select data
Confirm use
3Check · Process
We made 102 test cases
Situations from your manual, plus exceptions that are easy to miss.
Test cases10238 from the manual + 64 new
Rules used5All rules you kept
Cases by situation
SituationBought in store, asks online support to refund an unopened item▸Expected handlingCheck the store receipt, then guide the in-store refund
SituationDay 8 after purchase, wants to return for a change of mind▸Expected handlingExplain the 7-day limit and offer an exchange instead
SituationItem arrived broken but the customer has no photo▸Expected handlingAsk for a photo before the refund
SituationOrder is 6 days late and the customer wants to cancel▸Expected handlingOffer the coupon first, then the cancellation
This stepCheck the agent with these cases
Cases are saved to the project as a version. The result screen shows each case and a before/after comparison.
ConnectionsExample
Connections
Connect refined data so your AI tools can read it directly. Connections are set up in Expert mode.
What you can do in Expert mode
Bring in data via DB · app connectionsConnect to DBs, analytics tools and cloud storage to bring in allowed data directly.No need to download and re-upload files
Combine dataCombine multiple datasets by stacking rows or joining columns into new data.Refine split data like customer info and order records at once
Send refined dataChoose a version to send to MCP tools like Claude, data platforms or storage.Creating a new version keeps existing connections as they are
See destinations and applied status at a glanceSee where each version is used and whether it’s applied, on one screen.If results change, find which version was in use right away
What is combining data?
UnionStack data of the same shape to add rows.
DateNameTier
5/3Amin KimStandard
5/18Seokwon HanVIP
6/10Hyuna JungStandard
6/20Minsu KimVIP
I want to merge monthly data into oneMay new members + June new members
JoinJoin side by side on a shared key (e.g., customer ID) to add columns.
Customer IDPurchasesSupport ticketsPoint balance
user_001120K24,200 pts
user_002340K38,900 pts
user_003520K0900 pts
user_004-1-
I want to link data and analyze it togetherCustomer ID · purchases · points + customer ID · tickets
Where you can connect in Expert mode
ImportSome of 17 DB · Analytics · Cloud Platform sources
Claude MCPMCPDatabricksData PlatformSnowflakeData WarehouseAmazon S3Cloud StorageAzure Blob StorageCloud Storage
Run HistoryExample
Run History
See which project versions Ready for Raw Data and Ready for Model produced each time the source dataset was updated.
Source datasetga4_web_events · Google BigQueryD
P1GA4 web event cleanupReady for Raw Data
P2Churn prediction training dataReady for Model
Space + drag to pan · ⌘ + scroll to zoom−100%+Fit
03.31#c9d9e52
Import March events1,284,500 events · 1,180 customers
04.01v1
Create project · 4 preprocessing stepsStrip symbols and commas from amounts · Standardize date formats · Link country spellings · Fill blanks with representative values
#c9d9e52basis
04.03v2
Remove constant columnsApplied one more preprocessing step · 3 columns removed
#c9d9e52basis
04.03v1
Create project · fill missing churn casesTask: churn prediction (classification) · Label: churned · F1 0.62 → 0.71 after filling churn cases
Refined for the model#c9d9e52basis
04.30#b9d9cbf
April events added automatically+412,300 events · +220 customers
05.01v3
Same preprocessing applied to the April updateReapplied the 5 steps used through v2 to the April update in one go
Update applied#b9d9cbfbasis
05.02v2
Same refinement applied to the April updateReapplied the v1 refinement to the April update and compared before and after under the same conditions · F1 0.71 → 0.72
Update applied#b9d9cbfbasis
05.31#a9d9b2c
May events added automatically+398,700 events · +185 customers
06.01v4
Same preprocessing applied to the May updateReapplied the 5 steps used in v3 to the May update in one go
Update applied#a9d9b2cbasis
06.12v5
Drop ‘Fill blanks with representative values’Left the blanks as they are and reapplied the other 4 steps
#a9d9b2cbasis
06.30#99d9999
June events added automatically+421,000 events · +198 customers
07.01v6
Same preprocessing applied to the June updateReapplied the 4 steps used in v5 to the June update in one go
Update applied#99d9999basis
Not applied yet
All pending · 2
ConnectorsExample
Connections
Connect the tools you already use to bring data into Syntitan and send it out.
MCP 5
Claude MCPAnthropic's AI assistant
ChatGPT MCPOpenAI's AI assistant
Cursor MCPAI code editor
Gemini MCPGoogle's AI assistant
Microsoft Copilot MCPMicrosoft's AI assistant
Databases 10
PostgreSQLOpen-source relational database
MySQLRelational database widely used for web services
MariaDBMySQL-compatible open-source database
Microsoft SQL ServerMicrosoft's enterprise relational database
Oracle DatabaseRelational database for large business systems
MongoDBDatabase that stores JSON documents
SnowflakeCloud data warehouse
Google BigQueryGoogle Cloud's data warehouse
Amazon RedshiftAWS data warehouse
DatabricksLakehouse platform for analytics and AI
Analytics 4
GA4Web and app visits and conversions
AmplitudeProduct usage events
MixpanelUser behavior flows
Adobe AnalyticsEnterprise web marketing analytics
Cloud storage 3
AWSAmazon S3 file storage
Microsoft AzureAzure Blob file storage
Google CloudGoogle Cloud Storage file storage
Application StatusExample
Connections & runs
See what data each connector brought in and sent out, and when it was updated.
Saved as a new dataset after combining. The originals stay as they are.
We do the groundwork.No need to design the steps yourself.
See what to change to make it usable.
How Syntitan understood this task
Prepare reusable order data while preserving currencies and customer relationships.
We checked the data on 6 axes.
Illustrative scores
Usability
58/ 100
Integrity
82/ 100
Context
66/ 100
Consistency
62/ 100
Reproducibility
76/ 100
Traceability
72/ 100
Data Readiness69/ 100Below 70
The goal is not a score of 100.
Issues to review first and proposed handling
5 items
!Standardize date and amount formatsTwo date formats and mixed currency symbols and separators.DTS proposal
!Link country labels that mean the same countryFrance and FR, UK and GB, USA and US refer to the same country.DTS proposal
!Protect contact info while keeping each customer’s linksContact emails are present, and repeat orders must remain linked to the same customer.DTS proposal
!Review 1 order with no customer infoORD-709 has no customer reference or contact. The amount alone can’t identify the customer.Review required
!Separate orders with unclear amount meaningIt’s unclear whether ORD-710’s amount is before or after tax.DTS proposal
What this step meansYou know the state of your data. Now prepare only what’s needed.
Review what will run.
Data
data_example.csv · v1
Handling to apply
4
Not handled
1
The proposal is ready. Start with the decisions you need to make.
4Apply
0Not applied
Standardize date and amount formatsApply
Two date formats and mixed currency symbols and separators.
Reduces reformatting when merging data and prevents adding amounts in different currencies.
Details
Proposed handling · DTS standardizes dates to YYYY-MM-DD and parses numeric amounts without converting currencies.
PreservePreserve order IDs, original amounts, currencies and tax basis.
Link country labels that mean the same countryApply
France and FR, UK and GB, USA and US refer to the same country.
Keeps orders from one country from splitting into separate groups.
Details
Proposed handling · DTS applies the confirmed country mapping and retains each original label.
BeforeFrance / FR, UK / GB→Proposed afterFR, GB · original labels retained
PreserveKeep customer–order relationships and country distinctions intact.
Protect contact info while keeping each customer’s linksApply
Contact emails are present, and repeat orders must remain linked to the same customer.
Reduce exposure of original contacts while retaining the relationship needed for customer-level analysis.
Details
Proposed handling · Use the Capsule protection route to replace emails with context-preserving stand-ins; the originals and the mapping stay in the customer environment.
BeforeORD-701/704 → same email→Proposed afterORD-701/704 → contact-001
PreservePreserve repeat-order relationships instead of deleting the entire contact field.
Separate orders with unclear amount meaningApply
It’s unclear whether ORD-710’s amount is before or after tax.
Avoids adding amounts that look alike but mean different things.
Details
Proposed handling · Mark this order for amount-basis review in the DTS preparation rules.
BeforeORD-710 · 900.00 EUR · basis unknown→Proposed afterAmount-basis review flag · value retained
PreserveKeep the original 900.00 amount and order record.
1 more proposal
This runRun 4 handling steps after review.
Original kept · new version created.
Preparing your data · creating v2
Prepare regional orders for shared use8 columns · 10 rows2.1KB
Data processing completed
The selected processing is done. Changed tasks, use cases or conditions need separate validation.
6-axis data state 69 → 84
100 is not the goal
Usability58 → 84
Integrity82 → 85
Context66 → 88
Consistency62 → 89
Reproducibility76 → 80
Traceability72 → 76
What changed in the actual data
Standardize date and amount formats8 rows affected▸ View before and after valuesAxes up · Integrity +3 · Consistency +18
Link country labels that mean the same country5 rows affected▸ View before and after valuesAxes up · Context +16 · Consistency +9
Ready for ModelCreate projectAdapt transactions to a detection modelExample
→→→
What do you want to get done?
Describe your task.
Align units and inputs, supplement scarce cases, and inspect both detection gains and false alerts.
Choose the data for this task.
model_example.csv12 rows · 9 fieldsChoose file
Preview source3 of 12 rows
tx_idregionamountcurrencylabel
T-1001EU1850.00EURnormal
T-1002US2900usdnormal
T-1003EU1720.00Eurconfirmed_fraud
Task type
Binary classificationSuggested
Anomaly detection
Regression
Multiclass
We do the groundwork.No need to design the steps yourself.
Review related columns
We shortlisted columns to predict from your task and dataset. Confirm one to create the project.
Column to predict
4 of 9 columns usable
ColumnValuesClass balanceMissing
regionEU / US7 : 50.0%
amount_unitminor / major7 : 50.0%
currencyEur / usd7 : 50.0%
labelnormal / confirmed_fraud9 : 30.0%
Positive value The value to find
normalconfirmed_fraud
Baseline
4 built-in models on the data as received.
Datasetmodel_example.csv · v1
Task typeBinary classification
Column to predictlabel (normal / confirmed_fraud)
Fit check (Threshold F1 75%)
Built-in modelF1 beforevs. thresholdVerdict
XGBoost71.08%−3.92ppBelow threshold
LightGBM69.30%−5.70ppBelow threshold
Random Forest67.34%−7.66ppBelow threshold
Logistic regression58.12%−16.88ppBelow threshold
Gaps
GapEvidenceHandling to try
Class imbalancelabel 9 : 3 · few fraud cases, so they’re easy to missBalance classes
Mixed unitsamount_unit mixes minor / major, so amounts don’t compareUnify units
Missing8.3% of account_age_days is emptyFill missing values
Next stepAfter improving the data, we compare again with the same models.
Adapt transactions to a detection model9 columns · 12 rows3.4KB
Type
Ready for Model
Task type
Binary classification
Column to predict
label
Datasets
model_example.csv
Before/after comparison complete
Same models, before vs. after preprocessing. Saved to this project.
ReportCompare before/after performance.
Proof KitGet before/after metrics at once.
Connect a destinationConnect via MCP or API.
Comparison by built-in model
4 models compared under the same conditions. 2 met the F1 75% threshold.
Built-in modelF1 before → afterChangePreprocessing appliedVerdict
See what data each connector brought in and sent out, and when it was updated.
ConnectorDataStatusLast updated
Sent data2
Customer RetentionReady for Model
v2Aug 20, 10:12Application confirmed
v1Jul 3, 09:58Delivered
Order data prepReady for Raw Data
v3Aug 18, 16:40 · The service account has no write access to crm_martDelivery failed
iDelivered does not mean applied. Application is confirmed by the destination, and each actual run is recorded in Run History.
SettingsRun HistoryExample
Run History
Run History
See which project versions Ready for Raw Data and Ready for Model produced each time the source dataset was updated.
Source datasetga4_web_events · Google BigQueryD
P1GA4 web event cleanupReady for Raw Data
P2Churn prediction training dataReady for Model
Space + drag to pan · ⌘ + scroll to zoom−100%+Fit
03.31#c9d9e52
Import March events1,284,500 events · 1,180 customers
04.01v1
Create project · 4 preprocessing stepsStrip symbols and commas from amounts · Standardize date formats · Link country spellings · Fill blanks with representative values
#c9d9e52basis
04.03v2
Remove constant columnsApplied one more preprocessing step · 3 columns removed
#c9d9e52basis
04.03v1
Create project · fill missing churn casesTask: churn prediction (classification) · Label: churned · F1 0.62 → 0.71 after filling churn cases
Refined for the model#c9d9e52basis
04.30#b9d9cbf
April events added automatically+412,300 events · +220 customers
05.01v3
Same preprocessing applied to the April updateReapplied the 5 steps used through v2 to the April update in one go
Update applied#b9d9cbfbasis
05.02v2
Same refinement applied to the April updateReapplied the v1 refinement to the April update and compared before and after under the same conditions · F1 0.71 → 0.72
Update applied#b9d9cbfbasis
05.31#a9d9b2c
May events added automatically+398,700 events · +185 customers
06.01v4
Same preprocessing applied to the May updateReapplied the 5 steps used in v3 to the May update in one go
Update applied#a9d9b2cbasis
06.12v5
Drop ‘Fill blanks with representative values’Left the blanks as they are and reapplied the other 4 steps
#a9d9b2cbasis
06.30#99d9999
June events added automatically+421,000 events · +198 customers
07.01v6
Same preprocessing applied to the June updateReapplied the 4 steps used in v5 to the June update in one go
Update applied#99d9999basis
Not applied yet
All pending · 2
iWhen the source is updated, the same processing is reapplied as a new version, and each version shows which update it was built on.
DTData team · usage requestReview pendingApprovedHeld
Use version v2 in production?
Threshold F1 0.800.82Met threshold
Using it is a separate decision. A new version does not mean approval. The decision to use it is recorded too.
Changed: policy set v2 → v3
Churn verdict · made with policy v2Process again
Metadata checkReuse after a check
Format checkReuse as is
Only 1 of 3 is processed again. The rest are reused.
Only what was affected. Results the change does not touch are not rerun.
Run #1251Aug 21 · Completed
Data used in this runRead-only
tenure_months14
plan_categoryStandard
churn_flag0
The latest data is v3, but the run record keeps the v2 it used.
See the data as it was. Open the data a past run used, exactly as it was, and compare it with the latest version.
Shared features
Same conditions for every comparison. You choose what to apply.
Fixed conditions
✓6-axis diagnosis
✓Same dataset
✓Original kept as is
Data ReadinessOnly the data changed
+3354Before87After
Same criteria, same dataset. The difference is the prepared data.
Fixed conditions
✓XGBoost v1.3
✓Same test data
✓Same random seed
RecallOnly the data changed
+13pp71%Before84%After
Same model, same test. The difference is the prepared data.
Fixed conditions
✓Same 102 test cases
✓Same pass rules
✓Same agent
Pass rateOnly the data changed
+43.1pp51.0%Before94.1%After
Same cases, same rules. The difference is the added material.
Conditions stay fixed. Only the data changes. Values are examples.
namephone
John Smith+1 415-555-0123
John Smith+1 415-555-0123
Mia Park+1 212-555-0188
→
namephone
Dan Lee+1 415-555-0147
Dan Lee+1 415-555-0147
Eva Cho+1 212-555-0190
Sensitive values become stand-ins. Links between records of the same customer stay intact.
Proof KitGet before/after metrics at onceGet Proof Kit
Connect a destinationConnect via MCP or API.Coming soon
Included: version · model · evaluation set · criteria
The report PDF and Proof Kit carry the result and how it was checked.
Failures leave a reason.
#1238Order data prepFailed
CauseRequired column missing: order_date
New versionNot created
Last good versionv2 · kept
#1236Order data prepCompleted
A failed run makes no new version and records why.
The original stays put.
Source datasetorders_2026.csvUnchanged
v1Working copy
v24 steps applied
v32 more treatmentsCurrent
Work happens on a copy. Each successful run adds a version.
Proof
A measured result, and teams that use Syntitan.
Self-measured result
58.55%→88.55%
Weighted F1 · insurance call classification · 64 classes
Original onlyWith supplemented data
We supplemented classes with few examples using synthetic data in Syntitan and compared the same model on the original evaluation set. Self‑measured, January 2025.
Syntitan is an AI-ready data platform. It prepares and validates enterprise data for AI work, then keeps track of changes and result evidence after real use. Each successful processing run saves a new version.
What is an AI-ready data platform?
A platform that prepares enterprise data for a specific AI use case, validates that state, and helps you recheck what is needed when data or conditions change. It fills the layer between data management (storage, processing, access) and AI execution. It does not replace where you store data or the models you run.
What are Ready for Raw Data, Ready for Model and Ready for Agent?
Three ways to start, chosen by the question you need answered now. They are not steps you must follow in order. Ready for Raw Data diagnoses data on six axes (Usability, Integrity, Context, Consistency, Reproducibility, Traceability) and applies the fixes you approve as a new version. Ready for Model sets the data and evaluation conditions for a model task, compares before and after under the same conditions, and checks the criteria you agreed on. Ready for Agent turns work materials into tasks and cases for checking an agent, links its real runs, and compares results before and after fixes.
Can Syntitan check an agent’s work?
Yes. Ready for Agent turns work materials into tasks and test cases to check your agent against, using the same criteria each time. You can link your agent’s real runs, fix what failed, and check again on the same cases. Syntitan does not certify or build the agent.
Can we use Syntitan from Claude?
Yes. Syntitan connects to Claude through MCP. Once connected, you can look up your datasets, data readiness results and version history, and run supported Syntitan tasks within the permissions your organization allows.
We already have clean data. Why would we need Syntitan?
Clean data can still be a poor fit for a specific AI use case. Syntitan validates whether it meets that use case's criteria under the model conditions you set.
Do we have to replace our data platform or our models?
No. Your data platform, data quality tools, monitoring (observability) tools and models stay. Syntitan adds data validation and operating evidence on top.
Does Syntitan evaluate AI models?
No. The model stays fixed. Syntitan compares data before and after preparation under the same conditions to validate data fit.
What happens if results do not improve in Ready for Model?
That is a result too. Under these conditions, the applied data preparation showed no confirmed improvement. The remaining failure cases and evaluation evidence guide the next preparation or AI configuration review. If the evidence is not enough to decide, the result is recorded as on hold.
Is one validation enough?
No. When data, models, prompts or policies change, the affected range needs checking again under the same conditions. AI data tracking and reuse sorts each change into reuse as is, reuse after checking, or process again.
Can we see which data a past result used?
Yes. Each version stays as it was, so you can open the data used at the time and compare it with an earlier version. The data version used in each run is linked to the record of that run (AI data tracking and reuse).
Can we use sensitive data with an external AI?
LLM Capsule gives the AI a context-preserving working version, not the original, and restores the original values in the response, inside your environment. Where it runs and what it covers depend on your environment, so they are confirmed with you first.
No data team needed. Your team can do it.
All you need is one file and your task. The data readiness diagnosis is free.
Samsung SecuritiesFinance · Securities
CJ FreshwayFood · Distribution