You will get an evaluation harness that measures and protects your LLM quality
Project details
Most teams ship LLM changes on vibes — a prompt 'feels better,' the headline number ticks up, and a rare but high-stakes slice quietly regresses. You hear about it from a customer.
I build evaluation harnesses that turn 'feels better' into evidence: a labeled test set from your task, scored with the right metrics, broken down by slice/intent — because an aggregate hides the failure that hurts. Then I wire it into a CI gate so every change is checked.
WHAT YOU GET
• A labeled eval set built from your task and data
• Right metrics: accuracy, recall/precision, MRR, faithfulness, or LLM-as-judge
• A stratified report — strong slices vs. failing ones, with examples
• A CI regression gate that blocks quality drops before they ship (Standard+)
• Reproducible scripts + a short Loom walkthrough
WHY ME
• 6 years building & evaluating conversational AI at a top insurer (BERT → LLM); 3 years in data mining.
• Open-source eval demos on my GitHub: stratified routing/tool accuracy + RAG recall/precision/MRR.
• 100% async — written updates, GitHub PRs, recorded walkthroughs. No meetings.
Tell me what 'good' means for your system and I'll tell you how to measure it.
I build evaluation harnesses that turn 'feels better' into evidence: a labeled test set from your task, scored with the right metrics, broken down by slice/intent — because an aggregate hides the failure that hurts. Then I wire it into a CI gate so every change is checked.
WHAT YOU GET
• A labeled eval set built from your task and data
• Right metrics: accuracy, recall/precision, MRR, faithfulness, or LLM-as-judge
• A stratified report — strong slices vs. failing ones, with examples
• A CI regression gate that blocks quality drops before they ship (Standard+)
• Reproducible scripts + a short Loom walkthrough
WHY ME
• 6 years building & evaluating conversational AI at a top insurer (BERT → LLM); 3 years in data mining.
• Open-source eval demos on my GitHub: stratified routing/tool accuracy + RAG recall/precision/MRR.
• 100% async — written updates, GitHub PRs, recorded walkthroughs. No meetings.
Tell me what 'good' means for your system and I'll tell you how to measure it.
Machine Learning Tools
BERT, ChatGPT, GPT-3, MLflow, NLTK, NumPy, pandas, Python, scikit-learnWhat's included
| Service Tiers |
Starter
$200
|
Standard
$600
|
Advanced
$1,300
|
|---|---|---|---|
| Delivery Time | 5 days | 10 days | 18 days |
Number of Revisions | 1 | 2 | 3 |
Model Validation/Testing | |||
Model Documentation | |||
Data Source Connectivity | - | ||
Source Code |
Frequently asked questions
About Zhibin
Enterprise AI Agent & Conversational AI Engineer | LLM, RAG, NLP
Macau, Macao - 7:29 pm local time
I help companies turn LLMs into reliable, production-grade AI agents and conversational systems - not demos, but systems that handle real users at scale.
For 6 years at Ping An Life Insurance (one of the world's largest insurers), I built conversational AI from the BERT era through large-language-model fine-tuning: intent understanding, dialogue management, and customer-service automation serving millions of users. Before that, 3 years at Ctrip (China's largest online travel platform) doing large-scale data mining, recommendation, and risk/fraud modeling.
What I can build for you:
- AI Agents & multi-agent systems - tool use, function calling, orchestration (LangChain / LangGraph)
- RAG pipelines & knowledge bases - vector DB, retrieval, grounding, hallucination control
- LLM fine-tuning & evaluation - SFT, LoRA/PEFT, domain adaptation, prompt engineering
- Conversational AI, chatbots & customer-service automation
- NLP, anti-fraud, knowledge graphs, recommendation & CTR models
Tech I use daily: Python, PyTorch, Hugging Face, LangChain / LangGraph, OpenAI & Claude APIs, BERT, vector databases (pgvector / FAISS / Milvus / Pinecone), SQL.
How I work: I confirm scope and acceptance criteria up front in writing, send clear written progress updates, and deliver clean code with documentation.
If you're building an AI agent, a chatbot, or a RAG system - or need an ML expert for fraud / recommendation - message me your goal and I'll reply with a concrete approach.
Steps for completing your project
After purchasing the project, send requirements so Zhibin can start the project.
Delivery time starts when Zhibin receives requirements from you.
Zhibin works on your project following the steps below.
Revisions may occur after the delivery date.
Agree on metrics & acceptance criteria
We define what 'good' means, the metrics, and the slices that matter most - in writing - before any scoring begins. No surprises later.