An end-to-end human-in-the-loop system for recovering missing YOLO labels with multiple single-class teacher models.

• Enumerates GT/AUTO relationships with IoU, IoS, center distance, area ratio, duplicate detection, and cross-class conflict evidence.
• Streams six teacher models sequentially to control GPU and host memory pressure.
• Provides a Vue 3 + Spring Boot + MySQL review platform with JWT/RBAC, project membership, image-level atomic claiming, renewable group leases, optimistic versions, and immutable audit events.
• Completed a real review round of 30,183 candidates and exports immutable decisions for dataset regeneration and 5090 retraining.
• Verified by Python tests, backend integration tests, frontend builds, operational runbooks, and real product screenshots.

Repository · Platform architecture · Completion case study

A Java 17 multi-module administration backend modernized from a learning project into an inspectable delivery case study.

• Replaced committed secrets with environment-driven configuration and documented the credential-rotation boundary.
• Added BCrypt with successful-login migration from legacy MD5, modern JJWT handling, request IDs, Actuator, Flyway, and OpenAPI.
• Added H2/MySQL-mode integration tests, Maven Wrapper, GitHub Actions, a non-root Docker image, and health-gated Compose services.
• Verified 7/7 tests with 0 failures; the committed Swagger screenshot is captured from the running application.
• Found and fixed a real converter-order bug that Base64-encoded /v3/api-docs, then protected the endpoint with a regression test.

Repository · Architecture decisions · Security notes

A reproducible Chinese semantic textual similarity service with both an explainable baseline and a BERT-LoRA backend.

• Audits symmetric pair leakage and label conflicts before evaluation, preserving a clean 1,343-pair test subset.
• Provides a calibrated TF-IDF character n-gram baseline and a BERT + LoRA model behind the same FastAPI/CLI service contract.
• BERT-LoRA reaches Pearson 0.8016 and Spearman 0.7933 on the clean split, versus 0.6369 and 0.6337 for the baseline.
• Includes dynamic padding, gradient accumulation, CUDA-only FP16, clipping, early stopping, model/data cards, Docker, CI, and six passing tests.
• The UI screenshot is captured from a real API call and displays backend selection, score, interpretation, and latency.

Repository · Data card · Model card

A sanitized FastMCP service and Codex skill that turn flow-level signals into bounded lookup, risk evidence, false-positive context, and triage reports.

• Keeps scoring deterministic while the agent handles explanation and workflow composition.
• Exposes five MCP tools plus a CLI from a standard installable Python package.
• Constrains language output to evidence, uncertainty, possible false positives, and next actions.
• Includes synthetic public data, English/Chinese skills, regression tests, and Python 3.10/3.12 CI.

Repository · Chinese design notes

A seeded ACO-GA grid path-planning reproduction upgraded with validity checks and regression tests.

• Fixed a crossover bug that silently reduced population size every generation.
• Replaced unbounded recursive map regeneration with bounded BFS-validated generation.
• Added headless CLI execution, geometric length, turn count, collision checks, endpoint checks, and CI.
• On the committed seed-42 case, GA reduced turns from 8 to 6; simplification reduced 19 waypoints to 7 while remaining collision-free.

Repository · Metrics snapshot

A preserved YOLOv5 single-class defect-detection experiment presented as an auditable case study rather than an overstated product.

• Retains real training curves, PR/F1 curves, confusion matrix, GT mosaics, and prediction mosaics.
• Separates project scripts from the external upstream YOLOv5 runtime and documents the compatibility boundary.
• Audits figure presence, dimensions, hashes, and runtime readiness in CI.
• Explicitly discusses why near-perfect validation plots do not prove field generalization without split manifests, leakage checks, negatives, and independent footage.

Repository · Experiment card

flowchart LR
    A[Define evidence and failure modes] --> B[Build the smallest reproducible core]
    B --> C[Test data, concurrency, and edge cases]
    C --> D[Add human review or agent tools]
    D --> E[Measure and preserve artifacts]
    E --> F[Document limits and deployment boundary]

I prefer projects that can answer five questions clearly:

1. What real problem was observed?
2. Which part is deterministic code and which part is model uncertainty?
3. How are failures, stale writes, duplicates, or false positives handled?
4. Which outputs can another engineer reproduce or audit?
5. What is still missing before production use?

• yolo11-luna16-demo: chest nodule detection demo with GUI and SAHI.
• unet-camvid-segmentation: road-scene semantic segmentation experiment.
• project11-vit-finetune: Vision Transformer fine-tuning on CIFAR-10.

• GitHub: @jiapengLi11
• Email: 2284238579@qq.com

Public repositories intentionally exclude private datasets, production traffic, checkpoints, credentials, and internal deployment material. Claims are limited to the evidence retained in each repository.