Agrannya Singh
Software Engineer · Backend, Distributed Systems & Vector Retrieval
Building reliable backend architectures, in-memory OLAP pipelines, and vector retrieval systems. Currently working on sequential recommendation models at Samsung PRISM R&D.
"An idiot admires complexity, a genius admires simplicity."
Professional Experience
Research and software engineering across backend systems, vector retrieval, and cloud infrastructure.
Built and trained a 3.5M+ parameter deep sequential model (PyTorch GRU with Self-Attention) on 50M+ music interactions (Yandex YAMBDA-50M), predicting next-track listening intent across 156k+ tracks.
Formulated a joint multi-task objective pairing contrastive learning with 4-class user intent classification, outperforming the 0.1655 semantic similarity baseline with GPU mixed precision (CUDA AMP).
Indexed 156k+ track embeddings using FAISS vector quantization (IVFPQ), achieving 90% RAM compression (1024B to 32B) while enabling sub-millisecond nearest-neighbor retrieval at scale.
Engineered an out-of-core DuckDB pipeline ingesting 156k+ tracks in 118 seconds under 8 GB RAM limits; built automated multi-epoch checkpointing engines for fault-tolerant model training.
Featured Projects
Vector search engines, recommendation pipelines, and game AI.
TuneTrace - Music Discovery Platform
AI / MLDeveloped a hybrid analytics core processing 3,000+ catalog rows by combining collaborative filtering matrices with semantic vector spaces using the all-MiniLM-L6-v2 transformer framework.
Built an NLP Context Builder generating 384-dimensional dense embeddings out of metadata tracks, enabling rapid high-precision vector match queries via pgvector utilizing Cosine Distance algorithms.
Packaged code via Docker to Azure Container Registry for deployment on Azure Web Apps; built an optimized dual-layer Redis caching strategy to intercept high-frequency relational read queries.
Engineered a 4-stage recommendation pipeline over 45k+ films combining BAAI bi-encoders (bge-base) for candidate search, Bayesian rating shrinkage, and BAAI cross-encoders (bge-reranker-v2-m3) for reranking; cut P95 latency by 90.6% to 1.58s with 0.784 NDCG@10 ranking relevance and zero hallucination.
Built an enterprise-ready Model Context Protocol (MCP) server over Server-Sent Events (SSE), enabling AI agents (e.g., Anthropic Claude) to query live catalog recommendations with a 100% tool completion rate across 50 stress-test scenarios while mitigating cloud reverse-proxy drops.
Led an engineering group of 5 to develop an autonomous game agent, pairing optimized Alpha-Beta Minimax pathing with a custom ResNet-like tunnel-based architecture for an 8x8 grid (OthelloNetV3) to lock a 50% win evaluation.
Integrated NASA CCMC DONKI feeds across 50,000+ historical records, providing real-time aggregation and visualization of solar flare and coronal mass ejection disturbance events.
Technical Stack
Backend services, OLAP data engines, vector search pipelines, and infrastructure.
Cloud & Infrastructure
Databases & Query Languages
AI & Machine Learning
Frameworks & Backend
Programming Languages
Terminal
Agrannya Singh - Terminal [v1.0.0]
Type 'help' for commands or 'projects' to view code repositories.
Education
Vellore Institute of Technology, Vellore
B.Tech in Computer Science & Engineering
Certifications & Honors
Google Cloud
Enterprise multimodal LLM orchestration and deployment on Vertex AI
NCERT Government of India
Secured State Rank 22 in stage 1 evaluation
MongoDB University
Vector indexing, semantic search, and RAG architectures
IBM
Foundation models and enterprise AI pipelines
Linux Foundation
High availability, container orchestration, and monitoring
IBM SkillsBuild
Virtualization, cloud security, and IaaS architecture
Get In Touch
Open for high-impact infrastructure, backend, and distributed systems roles.
Send a Direct Note
Messages are saved directly to the database.