Skip to content
Sai Eswar Gujjala
  • Professional work
  • Deployed internally

Supplier Audit Video Intelligence

A multi-modal pipeline that turns hours of supplier audit recordings into a small set of distinct, searchable keyframes with defect summaries.

My role
End-to-end owner.
Key result
About 72% fewer frames after deduplication.
Technologies
  • Python
  • OpenCV (SSIM, optical flow, Canny edges)
  • ResNet-18
  • CLIP
  • DBSCAN
  • GPT-4V
  • spaCy (named-entity recognition)
  • FAISS
  • AWS
Skills shown here

What I owned

  1. Problem framing
  2. Data preparation
  3. Modeling
  4. Evaluation
  5. Application/API
  6. Deployment

Built at Zetwerk for internal use. Described in simplified form; no customer or supplier names, data, code or internal links are shown.

In [1]:Problem and constraints

Supplier audits are recorded on video. Reviewers had to watch full recordings to find defects and evidence, which was slow, and the footage was hard to search later.

In [2]:Data and preparation

500+ supplier audit recordings. The footage is never shown.

In [3]:Baseline and evaluation design

How many frames remain after deduplication, and manual review effort compared with reviewing the recordings by hand.

In [4]:Architecture and modeling decisions

Cheap classic computer vision first; expensive models only on the few frames that are left.

  1. Keyframes: SSIM finds scene changes, and optical flow skips frames while the camera moves.
  2. Noise filter: edge density (Canny) drops blurry or empty frames, and a ResNet-18 model removes frames without useful content.
  3. Deduplication: CLIP embeddings are clustered with DBSCAN, keeping one frame per cluster (about 72% fewer frames).
  4. Summaries: GPT-4V writes a short defect-context summary for each keyframe.
  5. Search: spaCy pulls out key entities, and FAISS makes audits searchable by meaning.

In [5]:Deployment and operation

Status: Deployed internally

Runs on AWS (internal).

Out[5]:Results and evidence

  • About 72% fewer frames after deduplication.
  • About 65% less manual review, compared with reviewing recordings by hand.
  • Audits can be searched by meaning.

In [6]:Limitations, failures, and lessons

  • Dark or shaky footage still produces weak keyframes, and the generated summaries are a starting point that reviewers confirm.