You will get an agentic workflow architecture and reliability review
Project details
AI agents often look reliable in a demo but fail under real traffic because tool calls, state transitions, retries, structured outputs, and human approvals are not validated together.
I will audit one agreed workflow end to end. I trace inputs, orchestration, model calls, tool and API interactions, memory or state, error handling, observability, latency, and cost. You receive reproducible findings, severity-ranked root causes, and an implementation-ready hardening plan.
Depending on the tier, I can also review a multi-agent architecture and design validation scenarios for handoffs, idempotency, retries, timeouts, fallbacks, and human-in-the-loop controls.
This service is model-provider agnostic and can assess LangGraph or custom Python/FastAPI orchestration using OpenAI, Anthropic, Google, open-weight, or mixed providers. I do not assume an older model stack.
Scope protection: delivery begins after requirements are complete and covers the agreed workflow(s) in the selected tier. New integrations, production deployment, or implementation outside that scope will be quoted separately.
I will audit one agreed workflow end to end. I trace inputs, orchestration, model calls, tool and API interactions, memory or state, error handling, observability, latency, and cost. You receive reproducible findings, severity-ranked root causes, and an implementation-ready hardening plan.
Depending on the tier, I can also review a multi-agent architecture and design validation scenarios for handoffs, idempotency, retries, timeouts, fallbacks, and human-in-the-loop controls.
This service is model-provider agnostic and can assess LangGraph or custom Python/FastAPI orchestration using OpenAI, Anthropic, Google, open-weight, or mixed providers. I do not assume an older model stack.
Scope protection: delivery begins after requirements are complete and covers the agreed workflow(s) in the selected tier. New integrations, production deployment, or implementation outside that scope will be quoted separately.
AI Algorithms
Large Language Model, Transformer ModelAI Applications
AI Chatbot, AIOps, Conversational AIAI Development Language
PythonAI Tools
Azure OpenAI, Hugging FaceAI Models
GPT-4, LLaMAWhat's included
| Service Tiers |
Starter
$250
|
Standard
$600
|
Advanced
$1,100
|
|---|---|---|---|
| Delivery Time | 4 days | 7 days | 10 days |
Number of Revisions | 1 | 2 | 2 |
AI Model Integration | - | - | - |
Batch Normalization | - | - | - |
Database Integration | - | - | - |
Detailed Code Comments | - | - | - |
Image Upscaling | - | - | - |
MLOps | - | - | - |
Model Deployment | - | - | - |
Model Documentation | - | - | - |
Model Monitoring | - | - | - |
Model Testing & Optimization | - | - | - |
Model Tuning | - | - | - |
Natural Language Processing | - | - | - |
NLP Tokenization | - | - | - |
Pre-Training | - | - | - |
Prompt Engineering | - | - | - |
Setup File | - | - | - |
Source Code | - | - | - |
Frequently asked questions
About Yu-Ting
AI Systems Engineer | Complex Data, Retrieval & Agentic Workflows
Taipei, Taiwan - 5:52 am local time
I also build the platform around these systems: PostgreSQL data models, APIs, background jobs, access control, usage metering, billing and payment flows, and failure recovery. Technologies such as FastAPI, Supabase, Stripe, OCR, and VLMs are implementation choices selected according to the project—not my professional positioning.
I have worked in AI engineering and research since 2021. Since January 2024, I have led product engineering and research at Apertis AI, where I build:
• Verbatim: multimodal ingestion for PDF, Office documents, and OCR; hybrid Qdrant and lexical retrieval; reranking; source citations; and evaluation.
• Production AI infrastructure: multi-model routing, provider failover, usage metering, cost controls, Stripe billing, authentication/RLS, and APIs.
• Reliable agent workflows: tool calling, structured outputs, validation, retries, observability, and human review.
• Open-source AI: merged contributions to LlamaIndex, Vercel AI SDK, Kilo Code, Lightning AI’s lit-llama, Lobe Icons, and CAG across Python and TypeScript ecosystems.
Core stack:
Python, FastAPI, Go, PostgreSQL, Supabase, Qdrant, BM25S, Redis, React, TypeScript, Next.js, Docker, and OpenAI/Anthropic APIs.
Best-fit projects:
• RAG quality, retrieval, citation, or evaluation improvements
• Document parsing, OCR, and multimodal ingestion
• AI agents, tool integrations, and backend reliability
• Complete AI SaaS features from API to UI
• LLM routing, metering, billing, and observability
Research credentials:
I am the first author of a CIKM 2022 paper and a recent preprint on retrieval bottlenecks in multi-hop document question answering. I have served on the LREC 2026 Scientific Committee and as a program committee member or reviewer for ROCLING, LREC-COLING, and COLING.
I work async-first, communicate through clear written updates, and provide reproducible deliverables. I am available up to 50 hours per week. Scheduled calls are possible between 9:00 PM and 12:00 AM Taiwan time (UTC+8).
Steps for completing your project
After purchasing the project, send requirements so Yu-Ting can start the project.
Delivery time starts when Yu-Ting receives requirements from you.
Yu-Ting works on your project following the steps below.
Revisions may occur after the delivery date.
Scope confirmation and evidence intake
I confirm the agreed workflow, access, success criteria, and reproducible failure cases before analysis.
Workflow tracing and reproduction
I trace model calls, tools, state, retries, approvals, latency, cost, and observability across the agreed path.