AI Practice
Products and tools built with AI — each one documents how I think, choose, and iterate with AI.
Flowtype Local
GitHub ↗A voice-to-prompt macOS app — turn casual spoken ideas into structured coding instructions in one tap
- 5Pipeline stages
- 0Python deps
- 300MBLocal 4-bit model
VSA · Voyager Super Agent
Enterprise Information Agent project and evaluation-methodology — from POC and design to an Agent benchmark system
- 3Eval dimensions
- 30+E2E eval set
- 100%Reachable entry
VLM Collision-Detection Eval Platform
An evaluation loop where VLM models act as "expert judges" for autonomous-driving collision detection — AI judging AI, with humans keeping final say
- 55,996Eval frames
- 0.812Recall
- 10Iterations
Internship AI Toolchain
Two self-built AI automation systems — a market-weekly Agent and an internal AI-news curation platform
- 2Self-built systems
- 20+Incremental iterations
- 0→1From zero to live
dsh-billing-tui
GitHub ↗Peak/off-peak billing TUI plugin for DeepSeek Harness — watch token usage live, compute cost, and print a persistent ASCII receipt
- LiveCost monitor
- peak/offTiered pricing
- 2TUI/Web modes
Open Source
Pull requests I opened and got merged into open-source projects
fix(memory): validate knowledge graph entries when loading from disk #4717
Fixed the MCP memory server loading knowledge-graph entries from disk without validation, which could ingest corrupted or invalid data. Memory fix merged into the official modelcontextprotocol/servers repo.
fix(docs): balance code fences in cookbook_template.mdx #7054
Fixed the unbalanced code fence in mem0’s docs template that broke rendering. Docs-only, review approved, merged into main as my first shipped contribution.
dsh-billing-tui —— 让每个 DeepSeek Harness 用户心里有数 #413
My billing plugin dsh-billing-tui tracks token usage live, prices it with DeepSeek’s official peak/off-peak rates, shows running cost in the TUI status line, and prints a whale ASCII receipt via /billing. It’s listed under “Dashboards & Session UX” in the community-curated DeepSeek Harness tool list, so DSH users can find it right in the catalog.
AI Thinking
Practical notes on Loop Engineering, Agent Benchmark, and AI Coding.
A Zero-to-One AI Coding SOP: From One Sentence to Verifiable Delivery
Don't just tell AI to write code — clarify, propose, plan, execute in stages, and prove correctness with tests, plus an 8-stage loop and an experiment framework
12 AI tools I've actually used
From drawing diagrams and driving a browser to querying data — a hands-on list of the AI tools I've run for real, split into general and on-network
Washing the "AI flavor" out of your frontend
Building pages with AI is no longer hard—the hard part is not making it obvious. Quantify the "AI flavor", then wash it out with three escalating methods
AI Coding Tools & Dev Habits
How I actually use AI in the terminal — the tools, the plans, and my current setup
I work in the terminal. All my AI coding agents live in the same terminal, and I orchestrate them with one self-contained toolchain.
Open-source tools I use
herdr
An agent-aware terminal multiplexer: run multiple coding agents side by side with live blocked / working / done status, so I know in one glance which one is waiting for me. Sessions detach/reattach, so nothing is lost on disconnect. One agent per pane, scheduled by status lights instead of polling.
dsh · DeepSeek Harness / dsh-billing-tui
DeepSeek's official Agent Harness — everything is a plugin. I mainly drive DeepSeek models through its TUI, and I wrote a dsh-billing-tui: it computes real-time token cost with peak/off-peak metering and shows it in the status bar.
Codex CLI
OpenAI's terminal coding agent. I use it as the main agent / decision-maker: making plans, breaking down tasks, and doing acceptance. Pure terminal, supports AGENTS.md, sandboxing and approval policies — good for a "think first, then act" role.
Claude Code
Anthropic's terminal coding agent. Mature Skills, CLAUDE.md and subagent ecosystem — I use it for exploration and for design-sensitive work like de-AI-ifying the frontend, or for cross-review with Codex.
ego lite
Lets agents drive a browser that reuses my logged-in state, running automation in their own Space without fighting me for tabs. I use it for real navigation, screenshots and form-filling when verifying UI.
Coding plans I've tried
I've used the higher-end Claude 20x and GPT 20x tiers — powerful, but at a price that's hard to justify. On a budget, my stable setup is two $20 tiers: ChatGPT Plus + the Ollama Pro coding plan.
My current setup
When a single plan's quota isn't enough, I split "strong reasoning" from "cheap execution":
- Heavy lifting: in one herdr pane, dsh-tui calls Ollama Cloud's DeepSeek V4 (official) for the bulk work — writing code, batch processing — low cost, fine to run long.
- Decisions & acceptance: another pane in the same herdr session runs Codex as the main agent — deciding, decomposing, reviewing and accepting. I only spend the expensive quota at the key checkpoints.
- Execution sub-agents: DeepSeek V4 Flash (official) acts as sub-agents for the small tasks I hand it, reporting back to Codex for acceptance.
In short: the terminal is home base, herdr is the control room, expensive models are the brain, and DeepSeek's cheap models are the hands — a tight budget pushed me into a cleaner "who thinks, who does" split.
About
Hi, I'm Yiheng Zhao — an MS student in Electronic Information at Xiamen University, researching LLM / RAG / Agent. I build agents with AI, and everything else with AI, distilling both the methodology and the shipped work here.
Education
- Xiamen University · Electronic Information · MS (985) — LLM / RAG / Agent, 2024–2027
- China University of Geosciences (Wuhan) · Automation · BS (211), 2020–2024
Research interests
Selected outputs
A Survey of LLM Technologies for Low-Altitude Operation Safety
Accepted at SESC 2026 · recommended to Systems Engineering — Theory & Practice
AcceptedLow-Altitude Economy Insight Platform · Industry-Policy Intelligent Analysis System V1.0
Software copyright registration
Reg. 2026SR0424325Constraint-Guided Neuro-Symbolic Causal Evidence Subgraph Retrieval for Safety Knowledge Reasoning
A constraint-guided neuro-symbolic method for causal evidence subgraph retrieval
Under reviewWhere my next paper is headed
For safety knowledge reasoning, a constraint-guided neuro-symbolic approach to causal evidence subgraph retrieval — retrieving, from a knowledge graph, only the causal evidence chains that actually support a conclusion, prioritizing explainability and evidence reliability over plain semantic recall. It follows the same principle I hold in agents: answering without sufficient evidence is a failure.
Stack
Languages / Platforms
Agent / RAG
Evals / Observability
For my full background, see the résumé PDF.