Vishal Tyagi

Applied ML · computer vision · automation

Models that grade 90k sheets — not another chatbot wrapper.

Deterministic CV over a learned bubble detector. Ambiguous marks go to review. Inventory image recognition. Scraping that drives a decision, not a demo notebook.

  1. Warp and persist — a bad scan does not redo the batch.
  2. Align to printed marks before any density call.
  3. Relative histogram, not a GPU detector.
  4. Ambiguous marks go to a human queue. Silent misgrade is worse.
OMR write-up →

Why this desk, not a generic portfolio

TypicalHere
Fine-tune a blog generatorOMR pipeline: warp → align → density → rules → queue
Unverified latency claimsReproducible scores, no GPU, institutional exam constraint
Prompt playgroundsRAG and Hopfield work documented as experiments, labeled as such

Proof for this desk