FinCredit Copilot
FinCredit Copilot
Next.js · FastAPI · RAG
Jorlin Shi
AI application and Agent engineer turning real-world workflows into AI systems people can actually use.
Coding with Jorlin. I build intelligent agents, knowledge retrieval, workflow automation, and reliable human-in-the-loop experiences.
AI APPLICATIONENGINEER
3PROJECTS SHIPPED
3APPLIED AI DOMAINS
3PUBLIC REPOSITORIES
AGENTS, RETRIEVAL, EVALUATIONPYTHON, FASTAPI, NEXT.JS, RAG
RECENT PROJECTS
Three end-to-end systems built around code security, sales operations, and financial decision support.
MY EXPERIENCE
A compact record of the problems I owned, the systems I built, and the evidence that the work held up.
Apr 2026 — Present
Independent AI Product Practice
AI Application & Agent Engineer
Owned the path from problem framing and architecture to interface design, evaluation, and reproducible delivery across three public AI systems.
- Built a resumable agent runtime with budgets, checkpoints, specialist review, and human validation.
- Implemented hybrid RAG with BGE-M3, BM25, RRF, grounded generation, and retrieval evaluation.
- Connected typed APIs, multi-turn state, CRM automation, observability, and product-facing interfaces.
Jul 2025 — Jan 2026
Simulation & Modeling
Simulation Modeling Engineer
Translated engineering requirements into executable models and validation workflows, building the systems mindset now applied to reliable AI products.
- Developed and validated C++ / Python integration interfaces for simulation workflows.
- Built Simulink / Stateflow models and supported verification against expected behavior.
- Worked across requirements, implementation, debugging, and delivery rather than treating modeling as an isolated artifact.
EDUCATION & RECOGNITION
Academic backgroundMaster’s degree Bachelor’s degree
EDUUniversity of Electronic Science and Technology of China
Project 985Project 211
Master’s in Electronic Information
Chengdu, China
Southwest Petroleum University
Double First-Class
B.Sc. in Electronic Information Science and Technology
Chengdu, China
Selected recognition
Outstanding Graduate Student
Graduate Academic Scholarship · Second Class
Provincial First Prize
China Undergraduate Mathematical Contest in Modeling
MCM Honorable Mention
Team captain
Research & Innovation
1 EI-indexed first-author paper · 1 invention patent application (second inventor)
CORE STACK
LLM & Context Engineering
Model APIs · prompting · structured output · tool calling
Agent Orchestration
LangGraph · runtime harness · state · checkpoints · human approval
RAG & Retrieval
BGE-M3 · BM25 · RRF · chunking · grounded generation
Backend APIs
Python · FastAPI · Pydantic · REST workflows
Data & State
PostgreSQL · SQLite · Redis Streams · conversation memory
Workflow Integration
n8n · CRM / Notion sync · webhooks · idempotency
Evaluation & Quality
pytest · retrieval metrics · eval sets · Ruff · mypy
Observability
OpenTelemetry · Prometheus · logs · traces · failure analysis
Product Frontend
TypeScript · React · Next.js · bilingual interaction design
Delivery
Docker · CI checks · reproducible local and public demos
BUILD NOTES
A reliable agent needs a runtime, not just a better prompt
A practical architecture for long-running agents: explicit state, token and time budgets, idempotent tools, checkpoints, retries, and validation gates. The key question is not whether an agent can finish once, but whether a failed run can be explained, resumed, and evaluated without repeating completed work.
Durable executionCheckpointHITL
Read case study
Hybrid retrieval is only useful when you can measure the retrieval
Why dense vectors and BM25 fail in different ways, how reciprocal rank fusion combines them without pretending their scores are comparable, and how Hit@k, MRR, nDCG, citation coverage, and answer faithfulness reveal different failure modes in a small production-minded RAG system.
BGE-M3BM25 + RRFRAG evaluation
Read case study
Designing multi-turn agents without duplicating messages or losing context
A conversation is not a single request. This note separates conversation IDs, message IDs, and idempotency keys; models a PostgreSQL-backed event history; controls context windows; and uses a transactional handoff so CRM or Notion writes remain replayable instead of becoming hidden side effects.
PostgreSQLIdempotencyMemory
Read case study
Where the LLM should stop in a high-stakes lending workflow
An engineering boundary for financial AI: use retrieval for policy evidence and language models for explanation, but keep DTI / LTV calculations, policy versions, approval permissions, and final decisions in deterministic and auditable components with explicit human checkpoints.
Rule engineEvidenceHuman approval
Read case study
LET’S WORK TOGETHER
Have a role, a project, or an awkward workflow that might benefit from AI? Send me a note.