Building local-first AI systems with API-first backends, reviewable workflows, evaluation loops, Bad Case tracking, and honest AI boundaries.
| Area | Showcase Signal |
|---|---|
| 🧠 AI Application Engineering | Deterministic workflows, RAG runs, reviewable AI outputs, and evaluation records |
| 🚗 Computer Vision | YOLOv8 detection, tracking artifacts, trajectory semantics, and video analytics |
| 🧾 Document Intelligence | OCR, classification, field extraction, evidence refs, and procurement walkthrough rules |
| 🧱 Backend Engineering | FastAPI services, typed schemas, SQLAlchemy models, Alembic migrations, and API-first flows |
| 🖥️ Frontend Workbench | React / TypeScript dashboards for review, reporting, bad cases, and local demos |
| 🧪 Evaluation Loops | Bad Case tracking, local metrics, regression fixtures, and report exports |
| 🔐 Privacy-aware Local Delivery | Local-first design, ignored runtime artifacts, Docker Compose, and clear repository safety rules |
I am a Master of Information Technology student at UNSW. My work focuses on turning AI, computer vision, RAG, and backend workflows into usable engineering systems: APIs, local dashboards, persisted records, evaluation loops, Bad Case review, report export, and honest project boundaries.
Current focus areas:
- AI Application Engineering
- Computer Vision and video analytics
- Backend / full-stack engineering
- RAG and deterministic agent workflows
- Financial document intelligence
- Smart traffic analytics
- Local-first and privacy-aware product design
| Area | Tools and Patterns |
|---|---|
| Backend | Python, FastAPI, Pydantic, SQLAlchemy, Alembic |
| Frontend | React, TypeScript, Vite, Ant Design where applicable, Streamlit |
| AI / CV | YOLOv8, PyTorch, Ultralytics, tracking, trajectory and event rules |
| RAG / Agent | Retrieval, grounded answers, deterministic workflows, evaluation fixtures |
| DevOps | Docker, Docker Compose, GitHub Actions, local deployment workflows |
| Data / Storage | SQLite, PostgreSQL, pgvector, local artifacts, CSV / JSON / report outputs |
Local-first AI career preparation workbench.
Built with FastAPI, React, TypeScript, SQLAlchemy, Alembic, SQLite, a PostgreSQL-compatible deployment shape, RAG, and deterministic agent workflows.
- Connects resume upload / parse / versioning, JD profiling, match scoring, project rewrite suggestions, interview practice, study plans, and application tracking.
- Includes RAG knowledge base workflows, grounded answer runs, deterministic agent runs, evaluation records, Bad Case review, and privacy controls.
- Designed around reviewable outputs rather than hidden prompt state or automatic resume changes.
- Boundary: production-readiness is in progress; it is not a live SaaS, not an automatic job application sender, and the default AI path is deterministic/local.
Evidence-first financial document audit and procurement walkthrough platform.
Built with FastAPI, React, TypeScript, Ant Design, PostgreSQL, pgvector, Alembic, OCR provider abstraction, RAG, and a deterministic rule engine.
- Covers task creation, document upload, OCR, document classification, field extraction, procurement walkthrough rules, RAG evidence retrieval, and report generation.
- Provides human review, field correction history, Review Center, Report Center, Bad Case tracking, Evaluation Center, RBAC/admin, and provider readiness boundaries.
- Keeps source evidence, citations, rule results, and review decisions visible for audit-style inspection.
- Boundary: non-production public/synthetic acceptance only; it is not a production audit system and does not provide professional audit, legal, compliance, investment, or regulatory advice.
Local smart traffic event detection and analysis platform.
Built with FastAPI, React, TypeScript, YOLOv8, a DeepSORT adapter, trajectory engine, SQLite, and Docker Compose.
- Supports video upload, analysis runs, YOLOv8 detection/tracking wrappers, trajectory semantics, configurable zones, event rules, alerts, manual review, Bad Cases, local evaluation, and PDF/JSON/CSV report export.
- Final local run generated 14 danger-zone intrusion events, 14 alerts, 1 bad case, 2 zone statistic records, and local event metrics.
- Event detection is rule-based over tracked trajectories, zones, and configured rules, not a dedicated traffic-law model.
- Boundary: local validation prototype; not law-enforcement-grade, not production-ready, not an official COCO / TrackEval benchmark, and pixel speed is not calibrated real-world speed.
Vehicle and pedestrian detection, tracking, local demo, API serving, and Docker delivery.
Built with YOLOv8 / Ultralytics, FastAPI, React, Streamlit, Docker, tracking artifacts, and CSV/JSON outputs for local detection, tracking, lightweight video analytics, API serving, and Docker delivery.
- Covers dataset validation, YOLOv8n / YOLOv8s / YOLOv8m experiments, image inference, video inference, tracking artifacts, React dashboard, Streamlit demo, FastAPI API, and Docker smoke checks.
- Recommended model: YOLOv8s. Demo image detection returned Truck and mini-truck boxes.
- Demo video task generated 21,988 detection rows and 21,988 tracking artifact rows for local inspection.
- Boundary: local technical validation / course project; not a production deployment, not an official COCO / TrackEval result, current demo tracking uses a synthetic tracker, and current demo run does not claim effective event / ROI / line-counting results.
| Project | Domain | Core Workflow | Main Tech | Evidence / Outputs | Boundary |
|---|---|---|---|---|---|
| 🎯 CareerAgent | Career preparation AI workbench | Resume, JD, match, rewrite, RAG, interview, study plan, applications | FastAPI, React, SQLAlchemy, Alembic, RAG | Match reports, grounded answer runs, agent runs, evaluations, Bad Cases | Production-readiness in progress; not live SaaS |
| 🧾 FinancialAuditAI | Financial document intelligence | Upload, OCR, classify, extract, rule audit, RAG, review, report | FastAPI, React, Ant Design, PostgreSQL, pgvector | Evidence refs, rule results, review records, reports, evaluations | Non-production public/synthetic acceptance |
| 🚦 SmartTraffic | Traffic video analytics | Video, detection, tracking, trajectory, zones, events, alerts, reports | FastAPI, React, YOLOv8, DeepSORT adapter, SQLite | 14 demo events, 14 alerts, Bad Case, PDF/JSON/CSV exports | Local validation prototype; not enforcement-grade |
| 🚗 YOLOv8 Detection | Computer vision detection | Dataset checks, model experiments, image/video inference, tracking artifacts | YOLOv8, FastAPI, React, Streamlit, Docker | Detection rows, tracking rows, API outputs, Docker smoke docs | Local demo/course project; not official benchmark |
- Local-first systems: reproducible workflows, ignored runtime artifacts, and clear safety rules.
- API-first backends: typed schemas, persisted records, and runnable local flows.
- Workbench UIs: dashboards that expose state, evidence, outputs, and failure cases.
- Evidence and review: citations, field sources, traffic event evidence, and model output checks.
- Evaluation loops: Bad Case tracking, local metrics, regression fixtures, and report exports.
- Honest AI boundaries: deterministic defaults, explicit provider paths, and no unsupported production claims.
- Docker/local delivery: local demo paths without committing models, datasets, databases, reports, secrets, or generated outputs.
- GitHub: @xulong-jia
- Email: xulongjia556@gmail.com
- Open to collaboration on AI application engineering, computer vision, backend systems, and local-first AI workflows.