Real-time analytics for an EU fintech
EU fintech, 50-200 FTE (k>=5 cohort)
- p99 latency
- 4.2s380ms
- monthly cost
- 3800 EUR1100 EUR
ProofLive operation · Audited
ProofWorkflows,in live operation,with an audit trail.
First the live systems we build and operate, then ten processes in BPMN, the way the platform runs inside existing systems: the right provider for the task, with a branch and a human at the decisions.
Live systems
What we build and operate
Real, running systems, not demos. The public projects can be checked live; for client systems we report numbers from real measurement, with the name protected by confidentiality.
01
Fintech clearing house
Since 2023 · live
A high-availability clearing platform, with AI in operations
We architect and operate the live platform of an FX and digital-asset clearing house: WildFly and Java on PostgreSQL and Redis, behind Traefik, for high-frequency trading and clearing. Production runs on AWS EKS, the lower environments on our own OKD/OpenShift cluster. Inside, a sandboxed RAG/agentic knowledge hub and a self-healing incident agent do the work, and AI reviews the code on every Pull Request before the human review.
99.9%+actual availability (99.5% SLA)
~100 h/weekengineer-hours recovered from the internal RAG, with 24 users
every PRAI code review before the human review
Production AIDevOps · KubernetesFintech
02
news.lavx.hu · our own, public
Live demo
A fully autonomous, multi-LLM news portal, with no human intervention
Public proof of the same multi-LLM orchestration we also use in closed enterprise work: from source to publication every step is automatic, on a multi-provider gateway with a model fallback chain. Next.js 15 and Payload CMS on PostgreSQL and Redis, with multi-layer duplicate filtering and an LLM-safeguard phase. Viewable live.
~100/daypublished articles, with no manual intervention
18 137unique visitors / 30 days (70 407 total)
03
AI code review in CI · open source
Publicly visible
The AI review catches real bugs, not review theater
The same automated Pull Request review we run in the live fintech CI also works on open-source projects: AI reviewers check every PR before the human review. It caught real, linkable bugs before they were merged: privilege bypass, a race condition, a secret leak and path traversal.
auth-bypasscaught before merge (bazarr #207)
deadlockrace condition, fixed (opensubtitles-scraper #11)
path traversalcritical, fixed (provider-catalog #62)
Case-study records
Disclosure-backed details
These are the legal-reviewed records behind the Proof archive. Each keeps the client frame, scope, method and outcome together, without inventing a brand name.
AI-driven customer support triage
EU SaaS company, 200-500 FTE (k>=5 cohort)
- first-response time
- 8h12min
- tickets auto-resolved
- 0%38%
The names of client systems are under confidentiality; we report the numbers from real measurement. The public projects can be checked live and in the source code.
Ten processes, in BPMN
How the platform works
Real flowcharts: a start and an end event, tasks, decision gates with branches. The card colour is the actor, the model card shows the provider.
Frontier model AFrontier model BEU-region modelOpen source modelDistilled localOCRTTSASRGitHubJiraConfluenceMS TeamsGrafanaSAP
graph TD
S(("Grafana alerts")):::ev --> A["Prometheus + Loki · metric, log"]:::sys
A --> B["Agent · trace, runbook, Kubernetes"]:::ai
B --> C["Model · root cause · distilled local"]:::modell
C --> D{"Is the action allowlisted?"}:::gate
D -->|yes| E["Agent · restart, scale, rollback"]:::ai
D -->|no| F["PagerDuty · human escalation"]:::ember
E --> Z(("Recovered, logged")):::ev
F --> Z
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
graph TD
S(("GitHub PR opened")):::ev --> A["GitHub Actions · webhook"]:::sys
A --> B["Model · SAST, secrets, dependencies · frontier model A"]:::modell
B --> C["Agent · weighted PR comment"]:::ai
C --> D{"Critical finding?"}:::gate
D -->|no| Z(("Ready to merge")):::ev
D -->|yes| E["GitHub · hard gate, merge blocked"]:::sys
E --> W(("Waiting for a fix")):::ev
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
graph TD
S(("PR change")):::ev --> A["Confluence + Swagger · API docs"]:::sys
A --> B["Model · docs vs implementation · EU-region model"]:::modell
B --> C["Model · code review · frontier model B"]:::modell
C --> D["CodeQL + Jira acceptance criteria"]:::sys
D --> E{"All green?"}:::gate
E -->|yes| Z(("Audited, mergeable")):::ev
E -->|no| F["GitHub · merge blocked"]:::sys
F --> W(("Waiting for a fix")):::ev
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
graph TD
S(("Commit / push / PR")):::ev --> A["GitHub + Jira + CI events"]:::sys
A --> B["Model · what changed · distilled local"]:::modell
B --> C["Agent · risk, expected completion"]:::ai
C --> D["Web board + MS Teams card"]:::sys
D --> E["Human · business visibility live"]:::ember
E --> Z(("Shared state picture")):::ev
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
graph TD
S(("Every morning")):::ev --> A["GitHub + Jira · overnight changes"]:::sys
A --> B["Model · condense the essentials · EU-region model"]:::modell
B --> C["Agent · blockers, priorities"]:::ai
C --> D["Model · spoken brief · TTS (local)"]:::modell
D --> E["MS Teams + Slack · text and audio"]:::sys
E --> Z(("Shorter standup")):::ev
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
graph TD
S(("Severe incident")):::ev --> A["Grafana + PagerDuty · correlation"]:::sys
A --> B["Agent · timeline, scope"]:::ai
B --> C["Model · root cause · frontier model A"]:::modell
C --> D["Model · fix in code · frontier model B"]:::modell
D --> E{"On-call approves?"}:::gate
E -->|yes| F["GitHub · fix PR + test"]:::sys
E -->|no| G["Human · manual fix"]:::ember
F --> H["Confluence postmortem + Jira actions"]:::sys
G --> H
H --> Z(("Closed")):::ev
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
graph TD
S(("Ticket or call")):::ev --> A["Model · voice call transcript · ASR"]:::modell
A --> B["Model · classification · EU-region / local"]:::modell
B --> C["Agent · Confluence KB + previous tickets"]:::ai
C --> D["Model · draft resolution · frontier model B"]:::modell
D --> E{"Simple case?"}:::gate
E -->|yes| F["Zendesk · auto-close"]:::sys
E -->|no| G["Human · escalation with context"]:::ember
F --> H["Jira · weekly product feedback"]:::sys
G --> H
H --> Z(("Faster response")):::ev
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
graph TD
S(("Contract / RFP email")):::ev --> A["Model · document ingest · OCR"]:::modell
A --> B["Agent · clause extraction"]:::ai
B --> C["Model · risk analysis · frontier model A"]:::modell
C --> D{"Risky clause?"}:::gate
D -->|yes| E["Human · legal review"]:::ember
D -->|no| F["Automatic approval"]:::sys
E --> G["Salesforce CRM + redline"]:::sys
F --> G
G --> Z(("Fast turnaround")):::ev
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
graph TD
S(("Release candidate")):::ev --> A["GitHub · changelog, migration"]:::sys
A --> B["Argo CD + Grafana · canary, metrics"]:::sys
B --> C["Model · anomaly · distilled local"]:::modell
C --> D{"Within SLO?"}:::gate
D -->|yes| E["GitHub · promotion"]:::sys
D -->|no| F["Automatic rollback"]:::sys
E --> G["Model · release notes · frontier model B"]:::modell
F --> G
G --> H["MS Teams · notify stakeholders"]:::sys
H --> Z(("Safe release")):::ev
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
graph TD
S(("Vendor invoice")):::ev --> A["Model · invoice OCR · OCR (local)"]:::modell
A --> B["Model · line-item extraction · distilled local"]:::modell
B --> C["SAP ERP · three-way match (PO, receipt, invoice)"]:::sys
C --> D{"Discrepancy or suspicion?"}:::gate
D -->|yes| E["Human · exception approval"]:::ember
D -->|no| F["Automatic posting"]:::sys
E --> G["SAP ERP · posting + audit trail"]:::sys
F --> G
G --> Z(("Fewer wrong payments")):::ev
classDef ev fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#9aa08f;
classDef sys fill:#171b13,stroke:#9AA38C,stroke-width:1.4px,color:#E9E4D6;
classDef ai fill:#171b13,stroke:#C68F44,stroke-width:1.4px,color:#E9E4D6;
classDef ember fill:#1a140f,stroke:#CE6E2C,stroke-width:1.6px,color:#E9E4D6;
classDef modell fill:#1b1d12,stroke:#DBB46B,stroke-width:1.6px,color:#DBB46B;
classDef gate fill:#13160f,stroke:#DBB46B,stroke-width:1.6px,color:#E9E4D6;
Let's talk
Tell us what you want to see live.
Within one working day we reply with a concrete next step. On the call you talk to an engineer, not a salesperson.