Chat, JSON, and evaluation
CV feedback, rubric scoring, tutor assistants, structured extraction, LLM-as-judge checks.
Keep prompts compact and cap outputs for demos.
Verified Azure scope
Use this matrix to choose techniques that are actually available for the June hackathon. It separates what works now, what requires tutor approval, and what is excluded from sample projects.
Verified
May 25, 2026
Region
Korea Central
Active AI resource
oai-cvbot-demo0525
Active deployment
gpt-4o-mini (2024-07-18, GlobalStandard)
Cost numbers are rough planning estimates, not invoices. Microsoft bills in USD and the school agreement, exchange rate, and free grants can change the final amount. USD retail prices are converted with a classroom estimate of USD 1 ~= KRW 1,400.
These techniques are safe defaults for student projects and the sample-template scaffolds.
CV feedback, rubric scoring, tutor assistants, structured extraction, LLM-as-judge checks.
Keep prompts compact and cap outputs for demos.
Triage agents, FAQ lookup, routing decisions, human escalation simulations.
Use simple local tools first; never let the model execute arbitrary commands.
Real-time tutor chat, interview practice, live feedback screens.
Stream only the final user-facing answer, not private reasoning or hidden state.
Document screenshots, form images, product photos, handwritten-note prototypes.
Use public or synthetic images; avoid passports, IDs, health records, and private student data.
Skill matching, semantic recommendations, small RAG demos, duplicate detection.
Ask a tutor before deployment. Use local vectors for the hackathon unless a shared deployment is announced.
Search and recommender demos without creating Azure AI Search.
Limit the corpus to small sample data; document how production would replace it.
These services are available in Korea Central, but they are not active resources for the event yet.
SpeechServices F0/S0 in Korea Central
Speech-to-text and text-to-speech when a team has a strong voice-first concept.
Requires a new Speech resource and tutor/admin approval. Do not use Azure OpenAI Whisper for this event.
ComputerVision F0/S1 in Korea Central
OCR-adjacent image analysis when gpt-4o-mini vision is not enough.
Prefer the existing gpt-4o-mini vision path unless a tutor provisions this explicitly.
FormRecognizer F0/S0 in Korea Central
PDF/page extraction, tables, form layouts, receipts, invoices, structured documents.
Page-based pricing can climb quickly. Use a 1-3 page sample set during the event.
Do not design sample tracks around these services for the June event.
Deployment test failed in Korea Central for the available SKUs.
Instead: Use text-first demos, or request a separate Azure AI Speech resource if voice is essential.
TTS was not available in the current catalog; the audio preview option is deprecated/provisioned and quota is 0.
Instead: Keep the Bilingual Voice Assistant as an Azure AI Speech variant, not an Azure OpenAI audio sample.
Provider is not registered and no Search service exists in the active resource groups.
Instead: Use the template local-vector scaffold plus embeddings. Treat AI Search as a production upgrade note.
Provider is not registered for this subscription.
Instead: Show evaluation and flow design in code or slides; do not require Azure ML resources.
Provider is not registered, and the checked availability does not match the Korea Central resource plan.
Instead: Build the chat UI directly in the template app or notebook.
The shared subscription is budget-controlled for the KOICA-TIU event.
Instead: Use assigned resources only; ask tutors for any exception before creating anything.
Use token-heavy services only when the demo needs them. Idle infrastructure is the easiest cost to avoid.
~$0.86 / ~KRW 1,200
Assumes 2,000 input tokens and 800 output tokens per call on gpt-4o-mini Data Zone pricing.
~$0.022 / ~KRW 31
text-embedding-3-small is cheap enough for small recommender/RAG demos; storage/search infrastructure is the real cost risk.
~$0.097 / ~KRW 136
Approximate 1 vCPU + 1 GiB active hour before free grants, requests, logging, and model calls.
~$2.42/day / ~KRW 3,400
Not approved for this hackathon by default. A month is roughly $74 before semantic-ranker usage.
| Service | Status | Unit | Estimate | Guidance |
|---|---|---|---|---|
| Azure OpenAI gpt-4o-mini | Allowed now | Input / output tokens | $0.165 input + $0.66 output per 1M tokens | Main model for all sample tracks. |
| Azure OpenAI text-embedding-3-small | Ops-enabled | Embedding tokens | $0.022 per 1M tokens | Use only if tutors enable the shared deployment. |
| Azure Container Apps | Allowed for assigned app slots | Active/idle vCPU, memory, requests | $0.0864 per active vCPU-hour + $0.0108 per active GiB-hour; requests $0.40 per 1M | Scale to zero; do not run background loops. |
| Azure Container Registry Basic | Active for the demo stack | Registry day + storage | $0.1666 per day + $0.10 per GB-month | Keep images small; delete old tags after the event. |
| Log Analytics | Active for the demo stack | Ingested logs + retention | $3.11 per GB ingestion + $0.14 per GB-month retention | Avoid verbose debug logging in loops. |
| Azure AI Speech | Possible, not active | Audio hours / TTS characters | $1.00 per STT audio hour; neural TTS about $15 per 1M characters | Tutor/admin approval only. |
| Azure AI Vision | Possible, not active | Image transactions | Common image-analysis meters are roughly $1.00-$1.50 per 1K transactions | Prefer gpt-4o-mini vision unless this is approved. |
| Azure AI Document Intelligence | Possible, not active | Pages | Read about $1.50 per 1K pages; prebuilt/layout often about $10 per 1K pages | Use a tiny page set; do not batch large PDFs. |
| Azure AI Search Basic | Excluded by default | Search unit hour + semantic ranker | $0.101 per hour Basic; semantic ranker can add about $1 per 1K queries after free allowance | Use local vectors instead. |
Chat completion, JSON mode, tool calls, streaming, and image input all returned successful responses on gpt-4o-mini.
text-embedding-3-small deployed temporarily, returned 1,536-dimension vectors, and was removed after the test.
Azure OpenAI Whisper deployment failed in Korea Central for both Standard and GlobalStandard SKUs.
Azure OpenAI TTS was not available in the model catalog for this resource; the audio preview quota is 0.
Azure AI Search, Azure Machine Learning, and Bot Service providers are not registered for this subscription.
The numbers on this page were checked with Azure CLI resource tests and Microsoft pricing references. Re-check before provisioning a paid service.
Start with the technologies marked Allowed now, then ask tutors before adding anything else.