Data Scientist @ Myntra · M.Tech CSE, IIIT Delhi · BlackboxNLP @ EMNLP 2026
stackdhruv.com · LinkedIn · Google Scholar · hello@stackdhruv.com
I build production ML systems — ranking models, multimodal pipelines, and computer vision applications. Currently at Myntra on the Homepage Merchandising / Ads Ranking team, and a graduate researcher at the MIDAS Lab, IIIT Delhi, where my thesis work on separating perception failures from reasoning failures in multimodal LLMs is published at BlackboxNLP (EMNLP 2026).
| Period | Role | Where |
|---|---|---|
| Jul 2026 – Present | Data Scientist | Myntra, Bengaluru |
| Jan 2026 – Jul 2026 | Data Scientist Intern | Myntra, Bengaluru |
| Jan 2025 – Present | Graduate Researcher (Thesis) | MIDAS Lab, IIIT Delhi |
| Jan – Jun 2024 | Jr. SDE Intern | Ishitva Robotic Systems |
Myntra — Ads Ranking
- Shipped a replacement ads-ranking model that delivered a statistically significant CTR lift in a live A/B test, after improving offline NDCG over both the incumbent production ranker and a heuristic baseline
- Extended the ranking objective from CTR alone to funnel-depth (CVR) metrics using deep multi-task ranking architectures
- Selected the production configuration through controlled ablations on loss function (pointwise binary-logistic → pairwise), boosting rounds, and auxiliary scalar features
- Built resumable, idempotent Databricks / PySpark pipelines for user and ad embedding generation, then migrated the entire embedding stack after diagnosing quality anomalies through EDA
- Authored the modular training-to-production codebase, MLflow-tracked experimentation through Triton runtime plan-file conversion
Production metrics for this work are covered by an employment confidentiality agreement and are deliberately omitted.
MIDAS Lab — M.Tech Thesis (M.Tech CGPA 8.08/10)
- Built a five-condition diagnostic framework that attributes multimodal LLM failures to perception, to reasoning, or to their interaction, holding the reasoning prompt fixed and varying only the visual grounding
- Ran ~140 inference runs across 6 multimodal LLMs and 5 physics/geometry benchmarks on a thread-pooled OpenRouter backbone with sliding-window rate limiting and idempotent checkpointing, so 8–12 hour jobs resume without recomputation
- Found that adding the diagram breaks up to 46.1% of problems the model already solves from text alone; LLM-judge evaluation validated against a four-annotator majority vote at Cohen's κ 0.74–0.84
- Separately, fine-tuned Qwen-2.5-VL-72B-Instruct on 5× H100-SXM for annotation-anomaly detection, served through a GPU-backed FastAPI + Celery async batch backend — 100+ req/s, 35% lower end-to-end latency
| Project | Highlights |
|---|---|
| Mental Health Meme Classification | Extends the M3H architecture with visual feature maps and Qwen-2.5-VL-7B semantic triplet extraction over OCR text; fine-tuned MentalBART · 65% / 63% Macro F1 on anxiety and depression · PyTorch Transformers Streamlit |
| Advanced ANPR & Face Recognition | YOLOv8-nano plate detection + fine-tuned ResNet-18 classifier (92% precision, 91% recall) with a Siamese few-shot face-recognition path for low illumination · National Finalist, KAVACH-2023 · YOLOv8 ResNet FastAPI React Native |
| Student Dropout Analysis | Government dashboard turning UDISE+ data into ML dropout-risk forecasts · R² 0.9976 on the EduDropX dataset · SSIP 2022 State Winner · Published IEEE I2CT 2024 · scikit-learn React MySQL |
| Microservices Benchmarking | DeathStarBench social network across a local Docker Swarm cluster and GKE, benchmarking three replica/node configurations with two-tier observability — Pixie eBPF for real-time tracing, Prometheus for fine-grained metrics · Kubernetes Docker Swarm Pixie |
- Lost in Perception: Isolating Perceptual and Reasoning Failures in Multimodal Physics and Geometry Reasoning · BlackboxNLP Workshop, co-located with EMNLP 2026 · Budapest, Oct 2026
- Unlocking Enigmatic Pathways: Empowering Student Dropout Analysis with ML · IEEE I2CT 2024 · IEEE Xplore
AWS Certified Machine Learning Engineer – Associate · issued Jul 2025
Microsoft Certified: Azure Data Scientist Associate · issued Jun 2024
ML / AI · PyTorch · scikit-learn · Hugging Face · vLLM · OpenCV Infra · FastAPI · Celery · Docker · Kubernetes · Databricks · PySpark · Triton Cloud · AWS · Azure · GCP Languages · Python · C++ · JavaScript · SQL