The academic record behind Sanket

Results & model lineage

The deployed jepa-asl-b3 (ASL-pretrained, CC BY 4.0 data) at accuracy test 0.5526 / unseen signer 0.5136 in a 5.2 MB model — from a 0.3318 baseline — beside jepa-x12, jepa-rr4 and the Phase C result.

1 · Runtime — in the browser

One frame's path from camera to answer: the pose worker, the 64-frame LandmarkRing, the two lists, worker messages, and where privacy is enforced.

2 · Training program

Corpora → JEPA pretraining → the fine-tune recipe (infusion, synthetic signers, rr) → four seeds → distillation → int8 → Lab bundle.

3 · Evaluation — the gate

The accuracy test over 599 held-out clips, why in-corpus is only a guardrail, mirror TTA, and why frozen probes mislead on this encoder.

4 · The JEPA objective

Masked latent prediction with an EMA target; Phase B's last-layer collapse versus Phase C's all-layer teacher, with the pilot ablation.

5 · Experiment infrastructure

Colab A100 + Mac bridge + tunnel + daemons + Drive checkpoints — and the three lessons: idle timeout, pool memory, the name collision.

Final review hub

The complete report (81 pages), transfer from ASL, the deployed jepa-asl-b3 — unseen signer 0.3318 → 0.5136 since Review 1 — sign by sign, and the honest reading.

Review 2 hub

The deliverable checklist, progress at a glance from Review 1 to now, the honest reading, and the report draft.

Algorithms & techniques

Eleven techniques as implemented — landmarks, ISL-JEPA, the fine-tune recipe, distillation, int8 export, calibration, the mirror pass, retrieval, shape match, the dataset system.

Intermediate results

Accuracy test 0.4591 → 0.5509, unseen signer 0.4818 → 0.5000; the lineage, seed tables, in-app numbers and the negative results kept.

Report (50%)

Chapter plan and status, the draft PDF, and the dataset review loop that gave +7.0 pp over four seeds.

Expected outcomes

Review 1 targets scored, four next levers under one rule, the semester-2 paper seeds, and SDG 10.

Review 1 hub

The deliverable checklist, progress at a glance across all three dates, and the printable handout.

Literature survey

The 2024–26 self-supervision turn: JEPA, skeleton and sign-specific SSL, and signer-independent evaluation — with the threads the experiments confirmed.

Novelty justification

The 0.82→0.33 cross-corpus measurement, shipped mirror TTA, ISL-JEPA — and the JEPA path now 15 pp above the supervised model.

Architectural design

The deployed four-worker runtime, the training program as run, and the evaluation contract binding them — proposed versus built.

Expected outcomes

Targets beside the measured column: Phase B's 0.50 gate missed at 0.3273, recovered to 0.4818; risks with what happened to them.

Review 0 hub

Who, what, when — the deliverable checklist, approval status, where the proposal stands now, and the printable handout.

Title identification

The chosen title, what each phrase commits us to and how each commitment has been kept, and the alternates considered.

One-page abstract

Problem, the two-lane on-device system, measured results, and plan — one page, with the review-0 numbers kept in a note.

Literature review

Datasets, video and pose-based recognition, low-resource methods, on-device inference — and the gap this work sits in, with status.

Novelty proposal

Five concrete claims, each with the evidence at review 0 and the evidence now.

Proposed system & timeline

The pipeline as deployed, each proposed item's status, the evaluation protocol, and the plan to end of November 2026.

Glossary

Every technical, sign-language, dataset, and institutional term used across the docs, each explained in a sentence or two. Dotted-underlined terms link here.