Gaurav Jain
builds AI that ships.

I trained a proprietary try-on model on SD 1.5 inpainting with self-attention, and fine-tuned FLUX with LoRAs for realism and consistency. It shipped as a full SaaS with APIs, a web app and a Shopify app, and I iterated directly on feedback from brands.

Open the demo app →

Problem

High-resolution photos hold huge amounts of detail and very small faces, which off-the-shelf detectors miss.

Approach

To build the dataset, I split each image into 16 overlapping tiles so every face was covered, and annotated them with an open-source detector. I then designed a custom architecture for tiny faces at high resolution. At inference, a 4-tile split kept detection fast and accurate.

Impact

Face-detection complaints dropped sharply, and users began recommending the product for its reliability.

Problem

Classify photos as "kiss", "almost kiss" or "no kiss". The moment fills a tiny region of a ~6K image, and the first model scored 24%, worse than random (33%).

Approach

I used face detections to find the region of interest, smart-cropped around it, and trained a CNN on the crops. Weighted loss and down-sampling handled the imbalance (15K kiss images against ~100K for each other class).

Impact

Accuracy rose to about 80%, and the preprocessing made the whole application about 10% faster.

About 95% of the data fell into roughly 2,700 of the 10,000 categories. I trained a model on that majority, testing bag-of-words, TF-IDF and Word2Vec, and handled the long tail separately. Careful data preparation with a basic bag-of-words model reached accuracy on par with the top BERT-based entries at the time.

View on GitHub →

Lightroom's temperature range runs from 2,000K to 50,000K, but my analysis showed that changes outside roughly 2,400–9,000K barely affect the image. I focused a regression model on that range: a transfer-learned CNN that takes the image plus its current temperature and tint. It made white-balance correction faster for photographers.

YOLO locates the subjects, and a scoring algorithm rates each image on where the subject sits relative to the rule-of-thirds grid. The scores helped photographers choose their best shots.

No dataset existed, so I interviewed developers across departments to learn what drives build time, then engineered features with pandas. A decision tree performed best and also revealed which features mattered most. The deployed model gave developers accurate build estimates.

Mask R-CNN and U-Net were too heavy, so I designed a U-Net-inspired CNN tuned for speed and size. A simple Flask app lets anyone upload an image and get the background removed in near real time.

View on GitHub →