Deepak Dhinwa

Analytics Intern

Scapia, Indian travel fintech, May 2026 to July 2026

Looked at how people actually use a travel card, then built the models that turn that behaviour into a retention or a matching decision.

• SigLIP2hotel image matching
• AirflowGPU inference on Databricks
• Churnscores for retention, by stage

• Usage data was read for how the card is actually used, including activity that looks like abuse rather than a trip.
• Churn scores are staged, so retention can act before someone has already left. Hotel photos are matched with fine-tuned SigLIP2, and that inference is scheduled on Databricks GPUs through Airflow.

SportMind

NBA analytics, January 2026 to February 2026

A table can rank scorers. A graph can walk from a player to a team, a ring, and an award, and a separate model can give a win chance without the chat inventing one.

• 10seasons, 2013-14 to 2022-23
• 6,246document chunks in chat
• 658player forecast models

• About 1,400 players and 12,800 games sit in Neo4j. Questions about stats, awards, and head-to-head are Cypher, not a guess from a paragraph.
• Chat uses hybrid search over the chunks, then a cross-encoder rerank. The win model is XGBoost with isotonic calibration, held out of that store on purpose.

• Neo4j
• Weaviate
• LlamaIndex
• XGBoost
• FastAPI
• React

Vehicle speed and plate reading

June 2025 to July 2025

Detection finds a box. Tracking is what makes that box the same car a moment later, so speed and a plate belong to a vehicle rather than to a flicker.

• ~15 msvehicle pass on an RTX 3050
• 0.95plate detector mAP50
• 50epochs on 222 plate crops

• YOLOv8n marks cars, buses, and trucks. SORT keeps an id by matching each new box to the previous track.
• The plate is a thin object inside a much larger box, so it has its own detector. TrOCR then reads the crop.

• YOLOv8
• SORT
• TrOCR
• FastAPI
• Docker
• Azure

Driving policy in MetaDrive

August 2025 to November 2025

A camera alone copies the easy parts of driving. The car’s own state covers the bends the image attends to poorly, and a second phase of learning is what keeps the car up.

• 81.5%lower MSE than camera-only
• 7.5×longer episodes after PPO
• 90kexpert steps copied first

• EfficientNet-B0 on a 128×128 frame, fused with a 19-number vehicle state. Behavioural cloning on the IDM expert cut error by 81.5% against a camera-only policy.
• Grad-CAM on the missed turns looked at sky and roadside. PPO, started from the fusion checkpoint, moved average survival from 299 steps to 2,256.

• EfficientNet
• behavioural cloning
• PPO
• Grad-CAM
• ONNX

Reconstruction and colorization

September 2025 to November 2025, with Prof. Biplab Banerjee

One frozen visual backbone, two jobs: put a masked photo back together, and paint colour, without training two networks or the whole encoder.

• 2 headsRGB repair and LAB colour
• ~90%fewer parameters trained
• LoRA 32rank that fits a 4 GB GPU

• DINOv2 ViT-S/14 stays frozen. It already understands patches. LoRA and the last transformer block are what adapt.
• A shared decoder splits into an RGB reconstruction head and a LAB colour head. Images are COCO at 224×224 with 25% of the patches masked.

• DINOv2
• LoRA
• PyTorch
• COCO

Airline passenger satisfaction

October 2025 to November 2025

The score is only useful if the missing delays and the repeated survey answers are handled before a model is allowed to see them.

• 0.987ROC-AUC, held-out flights
• 0.935F1 on that same split
• 0.93R² filling arrival delay

• 103,904 passenger records. Arrival delay is filled from departure delay, not with a column average, because a typical delay would invent a flight that did not happen.
• After the survey items stop repeating each other, XGBoost reaches ROC-AUC 0.987 and F1 0.935. SHAP reads the score back as cabin and reason for travel.

• XGBoost
• SHAP
• PCA
• imputation