I build AI systems that go from research to production — not just notebooks.

At Publicis Sapient, I designed Bodhi-Atomize, a multimodal AI pipeline that decomposes ad creatives (images, video, GIFs) into structured JSON using Gemini 2.5 Pro, YOLO, and PaddleOCR. What used to take hours now runs in ~1.5 minutes — achieving 84.5% consistency and 89% correctness evaluated with DeepEval.

Previously at Lincode Vision Labs, I shipped CV models to production — RF-DETR at 1.8x faster inference than YOLOv8, and defect detection from 55% → 82% mAP through synthetic data and multi-stage training.

My research: FedFV-CV — a federated learning framework for finger vein biometrics achieving 1.21% EER and 98.48% TAR, outperforming standard benchmarks across 122,600 images.

Multimodal AI system decomposing image, video & GIF ad creatives into structured JSON for competitive marketing intelligence.

Gemini 2.5 Pro · YOLO · PaddleOCR
FastAPI · Redis · Celery · K8s + KEDA
DeepEval · Pydantic structured outputs

84.5% consistency · 89% correctness · Hours → ~1.5 min

Federated deep learning for finger vein biometric auth — privacy-preserving with custom FedWPR aggregation.

MobileNetV2 · Custom FedWPR algorithm
122,600 images · 5 federated clients
Random sampling · non-IID distribution

1.21% EER · 98.48% TAR@FAR=0.01 · Outperforms SoTA

AI-powered Slack bot with RAG semantic search over internal docs. Event-driven, real-time query processing.

LlamaIndex · ChromaDB · OpenAI embeddings
FastAPI · n8n automation · Slack API
Channels + DMs · Event-driven architecture

20+ docs indexed · 40% faster queries · 50+ daily queries

End-to-end RAG chatbot deployed on AWS — fully containerized with automated CI/CD pipelines.

LangChain · FAISS · FastAPI
Docker · AWS ECR · App Runner
GitHub Actions CI/CD pipeline

Production deployment · Fully automated CI/CD