DeepSeek — LLM Integration Expert

DeepSeek is the most cost-efficient frontier LLM available today. V3 delivers GPT-4-class output at roughly 1/30th the cost — $0.27 per million input tokens. The API is OpenAI-compatible, which means if you have already built on GPT-4, you can switch by changing three lines.

But the real leverage is what you can build when LLM calls cost almost nothing. Agents that run hourly. Batch pipelines over thousands of records. Research loops that call the LLM 50 times per task. None of that is economically viable on GPT-4 pricing. On DeepSeek it is routine.

This skill covers everything you need to run DeepSeek in production: → Model selection between V3 (fast, cheap, general) and R1 (chain-of-thought reasoning) → Streaming implementation — real-time output with interruption handling → Tool use orchestration — define tools, run the loop, validate inputs, cap iterations → Multi-turn conversation management — history, context pruning, prompt caching → Cost tracking — log token usage on every call, estimate monthly spend → Error handling — rate limits, timeouts, exponential backoff → V3 vs R1 routing — when to use each, how to route by task complexity

Varconi runs its entire agent stack on DeepSeek V3. Scott, Otto, Penny, and Radar all use this integration. This skill is extracted directly from that production usage.

Free. No strings. Install it, use it, ship faster.