How TezMakale runs on ozDNA

TezMakale is a live vertical AI product serving Turkish students — AI detection, detailed analysis reports, and token-based usage economics — running on ozDNA infrastructure. This is what production AI looks like before the B2B layer opens.

Challenge

TezMakale operates in a high-pressure segment: students under deadline need fast, reliable AI analysis — detection to verify originality, segment-level reports with cited evidence, and clear usage limits that map to real infrastructure cost. The live product is at tezmakale.com.

The core challenge was not building demos. It was shipping production AI where every workflow burns tokens, every user session has margin implications, and peak traffic (exam seasons) can spike inference cost overnight.

• Multiple AI workflows per session — detect, analyze, report — each with different model requirements
• Token-based pricing that must reflect actual LLM spend, not arbitrary credits
• Turkish-language academic calibration — generic models underperform on local academic prose
• Need for cost visibility before scaling marketing spend

Architecture

TezMakale connects through the ozDNA Gateway — a unified AI gateway that handles authentication, rate limits, and audit logging before any inference reaches the model router or RAG layer. The public-facing product lives at TezMakale (tezmakale.com).

The model router sends classification tasks to lightweight models and reserves premium models only when analysis quality thresholds require it. The RAG layer provides academic-mode calibration — retrieval hooks tuned for Turkish academic writing patterns.

Solution

ozDNA deployed as the full vertical AI infrastructure layer — not just an API wrapper:

• AI Gateway — single entry point for all TezMakale AI workloads with per-user rate limits and session tracking
• Intelligent Routing — detect workflows routed to cost-efficient models; premium models only when segment analysis requires higher fidelity
• Cost Optimization — per-workflow cost attribution visible in dashboard; automatic downgrade rules during off-peak
• Usage Analytics — token burn mapped to user credits in real time, protecting gross margin
• Prompt Registry — versioned academic prompts with rollback capability for A/B testing detection accuracy

AI Workflow

A typical TezMakale session flows through three instrumented steps:

1. Detect — User submits text. Gateway authenticates, rate-limits, and routes to the detect workflow. Model router selects the academic-calibrated classifier. RAG layer provides segment-level context. Response includes AI probability, confidence, and segment map.
2. Report — Detailed analysis with cited evidence sentences. Router selects cost-efficient models per section. Cost engine tracks margin per report tier.
3. Analyze — Usage analytics records token burn and cost. Audit log captures prompt hash and model ID for compliance.

// Typical workflow cost attribution
{
  "workflow_id": "tezmakale-detect-v3",
  "user_id": "usr_...",
  "steps": [
    { "step": "detect", "model": "gpt-4o-mini", "cost_usd": 0.002 },
    { "step": "report", "model": "gpt-4o-mini", "cost_usd": 0.006 }
  ],
  "total_cost_usd": 0.008,
  "credits_charged": 4
}

Production Benefits

Running on ozDNA gave TezMakale production-grade infrastructure without building an internal AI platform team:

• Margin protection — token credits map to real workflow cost, not guesswork
• Operational visibility — finance and engineering share the same cost-per-workflow dashboard
• Scale readiness — peak exam-season traffic handled by gateway rate limits and model routing, not emergency model swaps
• Investor-ready proof — live production metrics demonstrate vertical AI infrastructure, not a thin wrapper