PORTFOLIO_GUIDE.md

Portfolio and interview guide

This guide turns the repository into evidence of engineering decisions rather than a list of scripts.

90-second project pitch

I worked on a multi-class safety-detection dataset where source datasets had been annotated for different tasks. In combined scenes, one class could be labeled while another visible class was missing. Those omissions become false negatives during training.

I built a conservative label-recovery pipeline around specialist single-class teacher models. It first audits the YOLO dataset without a GPU, then loads one teacher at a time, streams inference in bounded batches, matches predictions only against same-class labels, and routes unmatched predictions to AUTO, REVIEW or IGNORE using class-specific thresholds. It never edits source labels.

The difficult part was operational reliability on a long K x N scan. I added transactional adaptive batching after CUDA OOM, atomic checkpoints, signature-validated resume, idempotent CSV/label writes, environment manifests and HTML evidence reports. The repository includes a synthetic public demo, tests and CI while excluding private data and weights.

Capability evidence

Capability	Repository evidence	What to explain
Data engineering	audit command	Schema validation, orphan labels, corrupt images and split leakage
Computer vision	IoU matching and class mapping	Why same-class GT matching differs from NMS deduplication
GPU engineering	Sequential teachers and adaptive batch	Peak memory versus total compute; safe OOM retries
Reliability	state.json and idempotent writes	Batch commit boundary and crash replay semantics
MLOps	doctor and manifest.json	Reproducing CUDA, package and configuration state
Human-in-the-loop ML	AUTO/REVIEW/IGNORE routing	Precision-first thresholds and audit evidence
Backend and concurrency	Spring Boot collaboration platform	Image-level pessimistic allocation, renewable group leases, optimistic versions, JWT/RBAC and project isolation
Evaluation and policy	calibrate command	Turning reviewed outcomes into precision/recall-constrained class policies
Ensemble policy	consensus command	Independent evidence, one-to-one matching and coverage/risk tradeoffs
Scalable similarity search	cluster command	Perceptual hashes, BK-tree radius search and conservative collision guards
Active learning	prioritize command	Uncertainty, dynamic class balance, visual diversity and biased-sampling boundaries
Software quality	package, CLI, tests and CI	Public fixture, privacy checks and release build
Communication	HTML reports and architecture docs	Turning model work into reviewable project evidence

Live demonstration

1. Run python examples/create_synthetic_dataset.py --output .demo-dataset.
2. Run yolo-label-recovery audit .demo-dataset --output-dir .demo-audit --hash-images --check-images.
3. Open .demo-audit/dataset_audit.html and point out the intentionally injected class error, orphan label and cross-split duplicate.
4. Open examples/demo_output/report.html to show the recovery quality report without exposing project data.
5. Open examples/calibration/output/calibration.html to explain why each class receives a different threshold.
6. Open examples/consensus/output/consensus.html to show unsupported AUTO candidates being downgraded.
7. Open examples/near_duplicates/output/near_duplicate_report.html to show review compression and split leakage.
8. Open examples/prioritization/output/prioritization_report.html to show a limited-budget six-class review queue.
9. Run yolo-label-recovery doctor to show environment diagnostics.
10. Open docs/assets/platform-dashboard.png and docs/assets/platform-review-multibox.png to explain the validated two-account web workflow and five-candidate grouped review without requiring a live server.

This demonstration works without a GPU or private model weights. A full teacher scan remains an optional second demonstration when suitable public weights and data are available.

Design questions to expect

Why not overwrite the labels directly?

Source immutability makes experiments reversible, reviewable and safe to compare. A derived label tree can be deleted without losing the annotation baseline.

Why six passes instead of loading all teachers?

The compute remains approximately K x N, but loading one model at a time bounds peak GPU memory. This is a deliberate throughput/reliability tradeoff.

Why are AUTO thresholds different by class?

Teacher calibration and object difficulty differ. Small smoking/slipper targets should not inherit a threshold justified by a large tractor detector. Thresholds are policy and need validation data.

How did you choose the thresholds instead of guessing them?

The calibrate command sorts reviewed candidates once and builds cumulative precision-recall curves. AUTO uses the lowest threshold whose Wilson precision lower bound reaches the target at the selected confidence level, subject to a minimum sample count. REVIEW uses the highest threshold retaining the target positive recall, which minimizes review workload. Unsupported targets produce no fallback threshold.

What makes resume safe?

The state advances only after the current batch's labels and CSV rows are flushed. If a crash occurs between output and state commit, replay checks stable candidate keys and exact label lines, preventing duplication.

Why not load both Teachers together and ensemble their tensors?

Separate scans preserve the one-model GPU memory bound and allow different detector implementations to participate through a stable CSV contract. The consensus stage uses one-to-one IoU matching, keeps evidence auditable and downgrades unsupported AUTO candidates instead of deleting them.

Does high confidence make a prediction ground truth?

No. It makes the prediction stronger evidence. The project retains REVIEW routing, visual samples and an audit trail because confidence alone does not prove correctness under domain shift.

Why not compare every image pair for near duplicates?

Brute force grows as O(N²). The cluster command stores compact fingerprints and uses a BK-tree for Hamming-radius candidate search, then applies aHash, aspect-ratio and low-texture luminance safeguards. It still treats every group as review evidence rather than an automatic deletion decision.

Can the active review queue be used to report model accuracy?

No. It deliberately oversamples uncertain, rare and visually diverse cases, so its acceptance rate is biased. It is a remediation queue for finding useful failures per reviewer-hour. Accuracy still requires a separate random or stratified-random audit sample.

Claims to avoid

• Do not claim the teachers eliminate all missing labels.
• Do not claim thresholds are universal.
• Do not quote private dataset metrics without permission and a precise test definition.
• Do not describe exact hashing as perceptual near-duplicate detection.
• Do not imply the synthetic demo is a production benchmark.

Honest scope makes the real engineering contributions more credible.