The loop
Every agent is a state graph: detect, diagnose, plan, fix, verify, report, with a retry edge from verify back to plan. It can be drawn, tested and capped.
Home / Platform
Platform
How the DigiQuench platform is built
One framework runs all 38 agents: a shared self-healing loop, a risk policy, guardrails and full telemetry, on a lean AWS setup that costs about $23 a month for the complete demo.
The moving lines show data and calls flowing between the services.
The loop every agent runs
INYour dataExports, databases and apps, read-only
01DetectRules and trained models find problems
02DiagnoseEach problem named, with evidence
03PlanRisk policy: fix now or ask you
04Fix + verifySafe fixes applied, then re-checked
05Approve + reportRisky items wait for one click; you get the report
Three frameworks, three jobs
Each tool does what it does best
LangGraph runs the loop, CrewAI writes, n8n conducts.
The writers
Eleven agents write listings, supplier emails, audit notes and ad variants. A three-role crew writes and checks the text; the reviewer rejects anything that breaks a rule.
The conductor
Schedules each agent, gives every agent a webhook, emails a person when a decision is needed, runs the nightly check and captures leads from this site.
AWS services
What each AWS service does here
A lean footprint: one server plus managed services only where they are cheaper or safer.
| Service | Role in the platform | Why this choice |
|---|---|---|
| EC2 (Graviton t4g) | Docker with the Agents API, Control Tower, read-only demo, n8n and 38 agents | Arm instances cost less for the same work; about $23 a month for the whole demo |
| Amazon Bedrock | Language models for diagnoses, reports and CrewAI writing (Amazon Nova, Anthropic Claude) | Pay per token, no API keys (IAM role), model choice per client |
| S3 + Glue + Athena | Durable bronze, silver and gold lakehouse; SQL over Parquet | Pennies a month for gigabytes; no database server |
| CloudFront + S3 | This website over HTTPS with security headers | Inside the always-free 1 TB a month |
| CloudWatch | Logs and six agent metrics with alarms | Embedded Metric Format, inside the free allowance |
| SES | Alerts, digests, supplier chasers, cart recovery | $0.10 per 1,000 emails |
| Lambda / AgentCore | Optional: agents on demand from EventBridge schedules | Scales to zero when clients need isolated runs |
Machine learning
Seven models where numbers matter
Language models write and explain. Decisions on numbers use small, explainable models.
| Model | Used by | Method | Demo metric |
|---|---|---|---|
| Fraud scoring | FraudShield | Gradient boosting + IsolationForest | PR-AUC 0.96 |
| Default risk | LoanProcess | Logistic regression | ROC-AUC 0.79 |
| Demand forecast | DemandCast, InventoryIQ | Gradient boosting with lag features | MAPE 12.8% |
| No-show risk | ApptIQ | Gradient boosting | ROC-AUC 0.61 |
| Provider claims risk | ClaimsSmart | Random forest on aggregates | Synthetic only |
| Customer segments | AudienceIQ | K-means (k = 4) | Silhouette 0.13 |
| Recommendations | RecoEngine | Item co-occurrence | Hit rate@5 0.47 |
Metrics come from demo data and prove the pipeline works, not production accuracy. We measure accuracy on your data during the audit.
Observability
A complete record of every run
The Control Tower shows people what happened; CloudWatch alarms tell you when something fails.
OverviewRunsApprovalsIssuesModelsLakehouseControl Tower
all 38 agents · last runAgent runs3838 succeeded
Issues found1,768263 planted
Fixed + verified1,703re-checked
Awaiting approval65high risk
Issues found vs fixed by industry
Most issues this run
16ReturnsBotheld
15PriceWatchok
25PatientFlowheld
30CareBotok
29LabLinkok
Step into DigiQuench.
Agents that heal your data.
Watch an agent find planted problems, fix the safe ones and wait for your approval on the rest. Thirty minutes, live.