Jun-seo Park

Articles by Jun-seo Park

• DeepSeek V4.1 Flash API: Clean and Tag 10,000 Comments, and What It Costs

  We pulled 907 real comments and ran DeepSeek V4.1 Flash API on 602. Output is 74-83% of the model bill. Short codes and off-peak hours cut it by 72-79%.

• Gemini Enterprise Agent Platform and MCP: a brand sentiment agent with Agent Registry and Agent Gateway

  Gemini Enterprise custom MCP accepts only OAuth or no auth. Here is the ADK, Agent Registry and Agent Gateway setup for a key-based MCP, with DJI tests.

• GPT-6.1 Sol API structured outputs: review tagging cost estimate

  Tag reviews with GPT-6.1 Sol strict JSON and the Batch API. Data pull: 412 live reviews. Model cost is an estimate at public prices: about $0.81 per 1,000.

• GPT-6 Intelligent UI for a Reddit and X brand report: MCP setup example and data tests

  GPT-6 Intelligent UI can answer with tables and charts. OpenAI does not say it draws MCP data. Our 3-brand report cost $0.003336 in API fees.

• GLM 5.3 in Claude Code: a Zhihu and Weibo topic script

  How to use GLM 5.3 in Claude Code with a Z.ai or BigModel key, then let it write a script that compares Zhihu answers and Weibo comments, with real costs.

• Gemini Deep Research API with a remote MCP server: social data you can check

  The Gemini Deep Research API accepts remote MCP servers but not custom functions. 7 live social data calls cost $0.0046. Setup, pricing, citation checks.

• DeepSeek V4 Claude Code Setup: Comment Cleaning Costs

  We ran the numbers for Claude Code on deepseek-flash, with EveryInfra MCP pulling 695 YouTube, Bilibili and Douyin comments. Peak price is 1/10 of Sonnet 5.5, but more than Haiku 5.5.

• Claude Code Subagents: Haiku Workers for TikTok Accounts

  One TikTok brand account is about 9,700 tokens of raw data. A Haiku 5.5 subagent returns a 95-token summary. 98 live calls, 429 handling and the cost for 50 accounts.

• Jev Ecosystem Observation (7): Content Scoring and Information Filtering from Startup Ideas to Persona Testing

  The second major ecosystem scenario: using Jev for content scoring, information stream filtering, and outreach performance prediction. We examine killmyidea, Mike Taylor's one-time Jev review experiment on 37 published documents, SuperX's 61-question virality model, and Crowdcheck's 10,000-persona simulation.

• Jev Ecosystem Observation (6): Code Quality Evaluation — Measuring Actual Numbers vs. Claims

  Dozens of ecosystem projects claim to use Jev for code quality evaluation, but only a few publish controlled experiment results, costs, and accuracy. We examine agentjournal's five findings, Supercov's mechanical lessons, and Jev Review's phased evaluation to determine appropriate and inappropriate use cases.

• Jev Ecosystem Observation (Part 5): Fueling Decision Primitives – How to Combine Live Data and Calibration

  Wrap-up of the first five parts of the series. The inputs for Noul, Choice, and Score are states, and where those states originate determines the upper limit of the entire pipeline: the division of labor between the data and judgment layers, three practical combination approaches, an executable public catalog as a starting point, and scenarios where combination should be avoided.

• Jev Ecosystem Observation (IV): The Four-Day Infrastructure Layer, Directories, Clones, and Platform Adoption

  An analysis of 500 repositories: approximately 10 official demonstrations, 15 platform integrations, 50 standalone applications, 90 agent tools, and a long tail of 300 SDKs. Directory economics, compatible clones, dual-gateway distribution, and platform-level adoption over four days of formation for a new infrastructure layer.

• Jev Ecosystem Observation (Part 3): Four Days of Failure Samples and When Not to Trust It

  Success stories can be intoxicating. A $1.43 rigorous evaluation, an official quickstart failure, and several product-disclosed flaws map out the trust boundaries of decision primitives: high confidence does not equal correct, dimension splitting does not equal better, and wide options will break.

• Jev Ecosystem Observation (Part 2): 30 Agent Guardrail Projects, What They Blocked and What They Let Through

  Supervising coding agents is the primary use case in the Jev ecosystem, with over 30 projects becoming homogeneous within 72 hours. We break them down into four guardrail architectures, examine real-world empirical examples, and distill reproducible design principles.

• Jev Ecosystem Observation (Part 1): What 220 Cases Tell Us About Decision Primitives in Four Days

  Four days after the release of TypeSafe AI's Jev, we used ten rounds of comprehensive scans to compile over 220 cases with primary sources. This is the first article in the series, covering three semantic primitives, six high-frequency patterns, recurring failure lessons, and the emerging ecosystem infrastructure layer.

• Calling Gemini with the OpenAI SDK: From Minimal Text Requests to Rollback-Ready Migration

  Isolate the compatibility boundaries of the SDK, base URL, and models. Validate non-streaming text requests, errors, and billing details before restoring parameters task by task, avoiding the assumption that client compatibility equals full service compatibility.