Case Study Review 2 · 27 September 2026 · 2:00–6:00 PM · SRM KTR campus, in person
Review 2 — implementation stage
The four deliverables the programme office asks for at this stage, in their order: the algorithms as implemented, the results they have produced so far, the report half-written, and what the final review should expect. Four weeks after Review 1 the model that runs in the browser is the same size and speed, and it is right on 55% of the accuracy test instead of 46%. The more important number is the one beside it — the unseen signer moved far less — and every page says so.
- Candidate
- Mithunish Prabhakaran
- Registration No.
- RA2512049015065
- Programme
- M.Tech Artificial Intelligence, SRMIST (2025–2027)
- Course
- 21CSC601T – Case Studies · School of Computing
- Main faculty (final review)
- Dr. Pitchaimanikam · pitchaib@srmist.edu.in
- Support faculty (Reviews 0–2)
- Dr. S Saravanan · saravans2@srmist.edu.in
- Review
- 27 Sep 2026 · 2:00–6:00 PM · SRM KTR campus, in person
- SDG alignment
- SDG 10 — Reduced Inequalities
- Contact
- pmithunish@gmail.com
- Final review (tentative)
- 11 Oct 2026, with the report
Deliverables
| # | Deliverable | Status | Where |
|---|---|---|---|
| 1 | Implementation of algorithms / techniques | Eleven techniques implemented and in use | Algorithms & techniques |
| 2 | Intermediate results obtained | Measured on the accuracy test; negative results kept | Intermediate results |
| 3 | Report completion (50%) | Draft chapters and status | Report |
| 4 | Expected outcomes | Review 1 targets scored; next levers ranked | Expected outcomes |
How to read this. Read left to right. The webcam frame becomes 107 landmark points and is then discarded; a 64-frame window of those points feeds three answers side by side — the 260-sign set (a probability), the full vocabulary (a similarity score) and shape match (no learned weights). The only thing the server does is hand out the static bundle.
Progress at a glance
| Thread | At Review 1 (30 Aug) | Now (27 Sep) |
|---|---|---|
| Unseen signer (RKMVU-220 top-1) | 0.33 supervised · 0.12 for a left-dominant signer without the mirror pass | 0.5000 jepa-x12 (int8 0.4909) |
| Accuracy test (599 clips) | not yet defined — a cross-corpus harness on RKMVU alone | 0.5509 jepa-x12 · 0.4591 for jepa-rr4, deployed since 4 Sep |
| CISLR test (379 clips never trained on) | — | 0.5805 |
| What the app shows (int8 + mirror pass) | — | 0.5546 jepa-x12 · 0.4538 jepa-rr4 |
| Model for the 260-sign set | supervised, 9.9 MB int8 | ISL-JEPA encoder distilled from twelve fine-tunes, 4.57 MB int8 |
| Training approach | ISL-JEPA designed; pilots passed their pre-registered gates | Phase B pretraining → fine-tune recipe → twelve-model ensemble → distillation |
| Data | 118,168-unit unlabelled pool | 123,340-clip pool; a reviewed dataset contract (v4 pack); 75 escalations raised, 35 settled |
| Evaluation | in-corpus vs cross-corpus on the deployed bytes | one accuracy test, ≥ 4 seeds per claim, calibration cross-fit, in-app parity |
| Practice mode | — | confidence floor 0.5 credits 69% of correct signs at 83% precision (jepa-rr4: 61% at 82%) |
| Report | — | 50% drafted — chapter status |
How these pages are drawn
The diagrams in this section are hand-drawn Excalidraw scenes — the same drawings the review deck presents, rendered here in the site's own colours and in either theme. Each has a "View larger" button. The architecture section keeps the Mermaid diagrams from 4 September; the drawings here describe the system as of Review 2.