Suture — Pitch Deck | Google Cloud Rapid Agent 2026

Google Cloud Rapid Agent 2026

SUTURE

Your data pipelines
heal themselves.

< 60s Break to green

7 Fivetran API methods

$2K Saved per incident

Live pipeline status

SF → BQ Salesforce → BigQuery (Opportunities) healed

HS → SF HubSpot → Snowflake (Contacts) healthy

ST → BQ Stripe → BigQuery (Charges) broken

ZD → SF Zendesk → Snowflake (Tickets) healing...

suture agent — live

$suture detect stripe_prod

schema drift: amount_cents missing

Gemini 3: amount → amount_cents (0.91)

healed in 1.4s

What broke last Tuesday at 3:12am

47 failed syncs.
One renamed field.

A Salesforce admin renamed revenue to annual_revenue.
Every downstream dashboard, ML model, and report went dark. Silently.

$2,000Cost per incident (engineering + downtime)

2–4 hrsManual fix time per data engineer

10,000+Companies using Fivetran today

The 3am scenario

Sarah Chen

Data Engineering Lead · Series B startup · 47 Fivetran connectors

3:12am — PagerDuty fires.
Salesforce admin renamed revenue → annual_revenue.
All 47 opportunity connectors are FAILED.

3:14am — Log into Fivetran. Find the failing connector.

3:28am — Read error logs. Identify the schema mismatch.

3:45am — Compare source vs destination schema. Column by column.

4:02am — Guess the mapping. Update config. Trigger re-sync.

4:38am — Confirm it worked. 47 more connectors to go.

Cost per incident

$2,000

Engineering time + data downtime + delayed decisions

2–4h

Manual fix time. Per incident.
Fully preventable.

Every Fivetran team has this problem

Salesforce API changes. Stripe deprecates endpoints. HubSpot adds properties. Schema drift is constant and inevitable.

The fix

Schema breaks.
Suture heals it.
Before you wake up.

Autonomous 5-stage pipeline: detect the failure, diff the schemas, let Gemini 3 reason through the rename, patch the connector, verify the sync is green.

< 60sBreak to verified green sync

Zero human intervention

100% mock-mode reproducible demo

Full reasoning trace via Arize Phoenix

suture.edycu.dev/dashboard

PIPELINE HEALTH MONITOR

AGENT ONLINE

12

Pipelines Healed

1.8s

Avg Resolution

94%

AI Confidence

0

Human Actions

SF Salesforce → BigQuery · Opportunities HEALED 1.8s ago

HS HubSpot → Snowflake · Contacts HEALTHY

ST Stripe → BigQuery · Charges HEALED 4.2s ago

ZD Zendesk → Snowflake · Tickets HEALTHY

Under the hood

5-Stage Healing Pipeline

FastAPI · Gemini 3 Pro · Fivetran REST API · Arize Phoenix · Supabase

Live demo

Break it. Watch it heal.

seed.py → break_schema.py → watch Suture → verify_demo.py

1

Pipeline Status Grid

SFSalesforce → BigQuery (Opportunities)BROKEN

HSHubSpot → Snowflake (Contacts)HEALTHY

STStripe → BigQuery (Charges)HEALTHY

ZDZendesk → Snowflake (Tickets)HEALTHY

Pipeline status — live refresh 5s

2

Schema Diff — Source vs Destination

SOURCE (drifted) DESTINATION (expected)

annual_revenue

→

revenue

arr_usd

→

arr

Gemini 3 confidence: 0.94 · Rename detected

Schema diff — source vs destination

3

Agent Activity Log

[03:12:01]INFOFivetran webhook received — connector_sf_001

[03:12:01]ALERTSync failed: schema mismatch at Opportunity

[03:12:02]INFOFetching schemas via Fivetran discover_schema()

[03:12:02]AIGemini 3: annual_revenue → revenue (0.94)

[03:12:03]PATCHCalling PATCH /connectors/sf_001/schemas

[03:12:03]DONEPipeline healed in 1.8s — status: HEALTHY

Agent log — Gemini 3 reasoning live

4

All Pipelines Healed

1.8s

Resolution Time

7

API Calls Made

94%

AI Confidence

0

Human Actions

SFSalesforce → BigQuery · HealedHEALTHY

Verified green — resolution time 1.8s

1Pipeline grid

2Schema diff

3Agent log + Gemini 3 reasoning

4Healed — verified green sync

What makes it work

Three technical edges.

Each one alone is unremarkable. Together they close a loop no one has closed before.

7 Fivetran API Methods

We don't just read from Fivetran — we write back. PATCH /schemas modifies the connector's column mapping programmatically. No other agent does this.

list · get · discover · modify · trigger · logs · state

Gemini 3 Semantic Reasoning

String matching can't tell if revenue was renamed or deleted. Gemini 3 reasons through context and returns a mapping with 0.94 confidence score.

Gemini 3 Pro via Vertex AI

Full Arize Phoenix Observability

Every healing decision is traced. Judges can inspect the full reasoning chain — what the agent detected, what Gemini considered, and exactly which API call fixed it.

OpenInference · spans across all 5 stages

The moat

Only Fivetran makes this possible.

Schema mutation API. Fivetran is the only data platform exposing granular column-level schema modification via REST. This is the core unlock.

10,000+ companies. Every Fivetran customer has experienced schema drift. The addressable market is the entire Fivetran install base.

Blue ocean. No competing product exists here. Schema healing is universally done manually, with spreadsheets and Slack pings at 3am.

Built in a weekend. Fivetran's API quality made a full healing pipeline achievable in 72 hours. That's proof of the platform's engineering quality.

Without Fivetran → 5 separate systems

Without Fivetran

With Fivetran

Custom connector registry

fivetran.list_connectors()

600+ source adapters

fivetran.discover_schema()

Schema mutation engine

fivetran.modify_schema()

Sync orchestrator

fivetran.trigger_sync()

Log aggregator

fivetran.get_sync_logs()

Take Fivetran out and you'd need months of engineering. With Fivetran, this is a 72-hour weekend build.

What shipped

Fully built.
Not a prototype.

SHIPPED IN 72 HOURS

Python tests passing190

Total tests (Python + TS)287

Fivetran API methods7 / 7

Break-to-green time< 60s

CI matrixPy 3.11–3.13

python -m pytest agent/tests/ -v

platform darwin -- Python 3.12 -- pytest-8.1.1

rootdir: /suture/agent plugins: asyncio, cov

tests/test_schemas.py::TestPipelineStatus::test_healthy_value PASSED

tests/test_detector.py::TestClassifyError::test_schema_mismatch PASSED

tests/test_fivetran_client.py::TestDiscoverSchema::test_annual_revenue PASSED

tests/test_diagnoser.py::TestComputeDiff::test_salesforce_scenario PASSED

tests/test_mapper.py::TestFormatReasoning::test_confidence_score PASSED

tests/test_patcher.py::TestBuildPatch::test_salesforce_scenario PASSED

tests/test_verifier.py::TestPollStatus::test_resolves_healthy PASSED

... 183 more tests ...

====== 190 passed in 1.49s ======

$ npm test -- dashboard/src/__tests__/

Test Suites: 12 passed, 12 total

Tests: 97 passed, 97 total

FastAPI agent

Next.js 16 dashboard

Cloud Run Dockerfile

Vercel deploy

Arize Phoenix traces

GitHub Actions CI

The ask

Help us make
3am pages
obsolete.

Fivetran Partner Track

7 API methods · Schema mutation · Deepest integration

$5,000

Grand Prize

Autonomous AI agent · Real infrastructure problem · Production-ready

$20,000

"The best data engineer is the one who never gets paged."

Live demo

suture.edycu.dev

Edy Cu

edy.cu@live.com

github.com/edycutjong/suture

Gemini 3 ProFivetran APIArize Phoenix FastAPINext.js 16SupabaseCloud Run