← All work

AI/ML platform for enterprise data

Clyra

A B2B data platform where teams ask their tables questions in plain language and ship models as APIs.

Role
Senior Product Designer
Client
Clyra
Platform
Web app · API & MCP
Team
Engineers, project manager

At a glance

  • −85%raw data to live API, from 2–4 months to about 2 weeks
  • 4steps: link, impute, evaluate, publish

Challenge

From raw tables to a model live as an API took two to four months of data engineering.

Approach

One guided workspace: link without shared keys, fill the gaps, evaluate, publish.

Result

About two weeks instead of 2–4 months: 85% less time.

01

The problem

Most of an organisation's most valuable data sits in tables, from spreadsheets and exports to databases, and only a few people can query it. Everyone else waits for an analyst, or decides without the data. In finance, healthcare and retail those tables also live in systems that were never meant to talk to each other: different IDs, different formats, plenty of gaps.

Before a model can help anyone, teams have to match records without a common key, decide how to treat missing values, prove the model is reliable and then hand it to engineering to deploy.

Positioning map · ML platforms for enterprise data

Category view from desk research. Positions are qualitative.

Fig. 01Positioning map of ML platforms for enterprise data

Research synthesis · 4 themes

01No shared keys“Patient IDs differ across EHR, billing and device data”“Records are matched by hand in spreadsheets”
02Gaps break models“Every analyst fills missing values differently”“No trace of what was filled, or why”
03Is it good enough?“Accuracy is reported once, then never checked”“Drift is noticed weeks too late”
04Stuck in the queue“Models wait on engineering to become APIs”“Business teams can't use them directly”
Fig. 02Where data teams lose time between raw data and deployment
02

My role

Role

Senior Product Designer

As Senior Product Designer I owned the experience end to end, working with engineers and a project manager: research, flows, the data-viz and design system, and the interactive prototype.

Worked with

  • Engineers
  • Project manager

What I owned

  1. 01Product & interaction design
  2. 02Information architecture
  3. 03Data-viz system
  4. 04Design system
  5. 05Prototype
Fig. 03Data scientists, analysts, operators, engineers and PM in one workflow
03

How I approached it

I broke the path from raw data to a live model into four steps people already recognised, then designed each one so it could be understood and overridden, not just automated.

Key decisions

  1. 01Link records by similarity, not shared IDsMost enterprise sources don't share a common key.
  2. 02Show how missing values are filled before trainingTeams trust imputed data when they can see it and change it.
  3. 03Held-out score and drift on the home screenModel health is the first thing operators check every day.
  4. 04Publish as an API or MCP endpoint in one stepRemoves the hand-off to engineering for deployment.
  5. 05Every answer shows where it comes fromAnswers cite the live endpoint and its reports; the LLM only narrates the predictions.
  6. 06Sensitive data stays governedScoped, expiring keys; patient data stays in its region and never trains shared models.
CLYRA · WIREFLOW · DESKTOP 1508 × 824WIREFLOW — Raw data to a live APIA data scientist connects messy sources, Clyra links and fills them, the model is scored on held-out data,then published as an endpoint that analysts can query in plain language.01Connect sourcesDSEHR, billing, CRM and device data,connected once and kept in sync.02Link without keysSYSTEMRows matched across tables with noshared ID, each link with a confidence.03Fill & configureDSImputed cells shown, not hidden;target and split picked in one step.04Held-out evaluationSYSTEMScored on unseen rows againsta baseline, before anything ships.05Publish endpointDSOne step from a run to a versionedendpoint, no engineering hand-off.06API & MCPDSKeys, scopes and code samples;agents can call it through MCP.07MonitorSYSTEMPredictions, held-out score anddrift, every morning on Overview.Beats the baselineon held-out rows?No · tune & re-runYeskeySTATE CHANGEEndpoint v1 live · API key issued ·every call logged for drift and audit.ASK BAR · FROM ANY PAGEAsk about readmit-risk-30d…⌘J opens chat on the selected endpoint.ask in plain languagePARALLEL LANEAnalysts & operatorsANALYSTSQuery the live endpoint in plain language,no notebook and no ticket to the data team.ASK CLYRA · QUESTIONWhich discharged patients are at highest risk?answerANSWER · TABLE OR CHARTPredictions come from theendpoint; the LLM onlynarrates them.LEGENDDS — data scientist's actionSYSTEM — automatic stepANALYSTS — ask & consumeDecision pointDEVICEDesktop web app · 1508 × 824 viewport · keyboard-first: ⌘K command bar, ⌘J chat.SYSTEM EVENTSSources sync → links proposed → gaps imputed → held-out score vs baseline → endpoint versioned → drift watched.
Fig. 04Wireflow: from raw data to a live API, with the analysts in a parallel lane

Design notes · keyless record linking

EHR

M. Rossi · 1958 · 10013

J. Chen · 1971 · 11201

A. Diaz · 1964 · 07302

98%91%64% · review

Billing

Rossi, Mario · 03/58 · NY 10013

Chen J. · 1971 · Brooklyn

Alma Díaz · 64 · Jersey City

  1. 1Match on several fields together, not on one missing ID
  2. 2Every link shows its confidence
  3. 3Low-confidence links go to a human review queue
Fig. 05Keyless record matching and confidence

Design notes · evaluation & drift

readmit-risk-30d · v3Live
Held-out score88.2%
Training score91.4%
!Data drift detectedPayer mix shifted to Medicare Advantage, +7 pts since v3Retrain
  1. 1Lead with the held-out score, not the training score
  2. 2Explain drift in plain language, with its cause
  3. 3One clear action to fix it
Fig. 06Model evaluation and drift alerts
04

What I built

A workspace where a model goes from connected data to a live endpoint, with an assistant that answers questions about predictions and data.

  • Keyless data linking
  • Missing-value imputation
  • Model evaluation & drift
  • One-click API & MCP endpoints
  • Ask Clyra assistant
Overview · sidebar collapsed
01Overview · sidebar collapsed
Data catalog & connections
02Data catalog & connections
Dataset profile
03Dataset profile
Quick run · sidebar collapsed
04Quick run · sidebar collapsed
Run result · held-out metrics
05Run result · held-out metrics
Endpoints · sidebar collapsed
06Endpoints · sidebar collapsed
Endpoint detail & versions
07Endpoint detail & versions
API & MCP
08API & MCP
Reports · sidebar collapsed
09Reports · sidebar collapsed
Fig. 07Hi-fi screens, from the daily overview to a live endpoint
Ask Clyra · question to chart
01Ask Clyra · question to chart
Fig. 08Ask Clyra, live: a question in plain language, an answer grounded in the endpoint, a chart

Typography · two families, one scale

AaSpace GroteskTitles · 500 600 700
AaInterInterface · 400 500 600
TitleGood evening.Space Grotesk · Bold 700 · 34
Section titlePredictions trendSpace Grotesk · Semibold 600 · 17
SubtitleHeld-out score on 181K unseen rowsInter · Regular 400 · 15
ParagraphYour models served 168.7K predictions today.Inter · Regular 400 · 14
KPI88.2%Inter · Semibold 600 · 30
LabelHeld-out scoreInter · Medium 500 · 10 · caps
Codereadmit-risk-30d · v4SF Mono · Regular 400 · 12

Color

Signal#F2622EInk#161616Stone#F6F3EFLine#E7E2DCHealthy#1E9E6A

Components

Publish endpointRun evaluation
LiveHealthyPausedDrift
Held-out score88.2%↑ 1.4

claims-integrity-score 67.5K

Search or run a command…⌘K

Chart the flagged claims by department Flags are concentrated: 2 departments account for 20 of the 38.

Radius

610121822
Fig. 09Type scale with weights, color and core components
05

The result

  • −85%raw data to live API, from 2–4 months to about 2 weeks
  • 4steps: link, impute, evaluate, publish

Getting a model from raw data to a live API used to take two to four months of hand-offs between data, ML and engineering. With Clyra the same path, linking data without common keys, filling missing values, evaluating and publishing, takes about two weeks.

Time from raw data to a live API

Before2–4 months · hand-offs between data, ML and engineering
With Clyra≈ 2 weeks −85%
01Link02Impute03Evaluate04Publish
Fig. 10Time from raw data to a live endpoint, before vs after

Contact

Let's build something people use and businesses need.

bressanleonardo@hotmail.comGet in touch