About and contact

1. Define the Spec First

   Before implementation, I write a numbered spec with acceptance criteria and a constitution: the working rules every agent must follow.

2. Delegate Implementation, Own the Review

   AI coding agents implement much of the code in parallel. I am responsible for design, review, verification in a deployed environment and the merge decision.

3. Evaluate Before Release

   I evaluate changes with gold sets, A/B tests, CI gates and shadow mode. I record negative results and leave changes disabled when they do not help.

4. Design for Recovery

   I plan for failures and build a recovery path rather than assume the system will always work.

5. Report Metrics with Context

   I publish a metric only when I can explain how it was measured: in the lab, offline or live.

• LLM telemetry with GenAI conventions. Each model call records token usage (including reasoning and cached tokens), model, provider and agent, so cost and latency can be traced per feature.

• Alerts as code. Alert rules, action groups and diagnostic settings live in Bicep next to the app and ship with azd: 44 rules in production today, covering daily and month-to-date spend, reasoning-token spikes, 429 throttling per deployment, judge latency and failed evaluation gates.

• Telemetry governance. Personal data is redacted before export. Noisy spans are sampled, and ingestion has a daily cap with an alert at 80 percent.

• Self-hosted when it fits. For a ticketing marketplace backend I provisioned Prometheus and Grafana as code, with black-box probes on the public API every 5 seconds plus Postgres, Redis and host metrics.

This site is also monitored: an anonymous performance beacon, nightly performance runs, and public visit stats at the Visitor Observatory in the Lobby.

Proof from production

• OpenTelemetry

  100%

  of surveyed AI services

• Alert rules as code

  44

  in production, defined in Bicep across 7 repos

• LLM token telemetry

  3

  production AI services

• Traced dependency calls

  ~1.2M

  per month on one queue-based scanner

• API probes

  10

  public endpoints every 5 s

param databaseAccountName string
param actionGroupId string

@description('429 responses per 15 minutes that fire the alert.')
param throttleThreshold int = 50

resource database 'Microsoft.DocumentDB/databaseAccounts@2025-10-15' existing = {
  name: databaseAccountName
}

resource throttling 'Microsoft.Insights/metricAlerts@2026-01-01' = {
  name: '${databaseAccountName}-throttling'
  location: 'global'
  properties: {
    description: 'Cosmos DB returned 429 TooManyRequests above the threshold in 15 minutes.'
    severity: 3
    enabled: true
    scopes: [database.id]
    evaluationFrequency: 'PT5M'
    windowSize: 'PT15M'
    autoMitigate: true
    criteria: {
      'odata.type': 'Microsoft.Azure.Monitor.SingleResourceMultipleMetricCriteria'
      allOf: [
        {
          criterionType: 'StaticThresholdCriterion'
          name: 'throttled'
          metricNamespace: 'Microsoft.DocumentDB/databaseAccounts'
          metricName: 'TotalRequests'
          dimensions: [{ name: 'StatusCode', operator: 'Include', values: ['429'] }]
          operator: 'GreaterThan'
          threshold: throttleThreshold
          timeAggregation: 'Count'
        }
      ]
    }
    actions: [{ actionGroupId: actionGroupId }]
  }
}

Related case studies

• AI Code Security PlatformPris.AI, MII's AI code security platform: eight scanners per pull request, readable AI triage and a fix agent that opens PRs with validation and guardrails.
• Realtime Voice AI Contact CenterARIA, MII's voice AI contact center: ~2 s to ~0.2 s interruption latency in a controlled A/B (n=4 calls per arm), grounded answers and human handover.
• AI Fraud Document ReviewMII Fraud Detection: OCR and two-pass LLM procurement review, with offline development-set evaluation. The hybrid engine remains in shadow mode only.
• Shared Azure AI OperationsModel migration in one day, coordinated key rotation, alerts managed as code and a Node.js security patch sweep across AI apps.
• Event Ticket Marketplace BackendA live ticket marketplace backend for auctions, resale, wallet accounting and partner settlement.

• AI and LLM

  • Microsoft Agent Framework
  • LangGraph
  • LangChain
  • Azure OpenAI (incl. Realtime)
  • Prompt Shields
  • Hugging Face
  • RAG
  • Hybrid/Vector Search
  • LLM evaluation

• Data science and ML

  • pandas
  • NumPy
  • Matplotlib
  • Seaborn
  • Plotly
  • Streamlit
  • Jupyter
  • scikit-learn
  • PyTorch
  • Hugging Face Transformers
  • Keras
  • Sastrawi

• Azure

  • Azure AI Foundry
  • Azure AI Search
  • Cosmos DB
  • Document Intelligence
  • Azure Container Apps
  • Azure App Service
  • Azure Functions
  • Service Bus
  • Key Vault
  • Managed Identity
  • Microsoft Fabric
  • Azure Monitor

• Languages

  • Python
  • TypeScript
  • JavaScript
  • Go
  • Java
  • PHP

• Backend

  • FastAPI
  • Spring Boot
  • Express.js
  • Gin
  • Echo
  • Laravel
  • CodeIgniter
  • Flask
  • Django
  • GraphQL

• Frontend

  • Next.js
  • React
  • Vue
  • Nuxt
  • Tailwind CSS

• Data and messaging

  • PostgreSQL
  • MySQL
  • MongoDB
  • Firebase
  • Redis
  • Kafka
  • RabbitMQ
  • Elasticsearch

• DevOps

  • Docker
  • Bicep
  • GitHub Actions
  • Jenkins
  • AWS
  • LiveKit
  • k6
  • SonarQube
  • Sentry

• View CV
• /en/cv