LinkedIn · Email · Ludhiana, Punjab, India

I build production-grade AI systems — not demos, not notebooks.

• 🔭 Currently working as an AI Research Associate, shipping GenAI systems in production
• ⚡ Cut LLM fine-tuning time by 51% (72h → 35h) through systems-level optimization
• 🧠 Improved model loss by 35% through systematic hyperparameter experimentation
• 🌐 Built multilingual OCR pipelines supporting 80+ Indian languages
• 🚀 Everything I build is measurable, deployed, and documented

Programming Python

ML & Deep Learning PyTorch ResNet LSTM Transfer Learning Time-Series Forecasting NLP Supervised/Unsupervised ML Recommender Systems

Generative AI LoRA Fine-Tuning RAG Architecture Agentic Workflows LangGraph LangChain OCR

MLOps & Infra FastAPI MLflow DagsHub Docker Docker Compose Apache Airflow Feast Feature Store Langfuse Evidently AI Prometheus Grafana

Data PostgreSQL Qdrant ClickHouse

Fully local, zero-vendor-lock-in research retrieval over arXiv CS.AI papers.

• LangGraph agentic workflow: guardrail validation → hybrid retrieval → LLM grading → adaptive query rewriting → grounded answer generation
• Hybrid search on OpenSearch: BM25 + 1024-dim Jina embeddings via Reciprocal Rank Fusion
• Daily Airflow DAG auto-ingests, parses PDFs with Docling, and upserts into the hybrid index
• 13-container production stack: FastAPI (REST + SSE), Gradio, Telegram bot, Langfuse v3 tracing, Redis caching
• 100% local inference via Ollama (Llama 3.2) — zero API cost, zero data leakage

Python LangGraph OpenSearch FastAPI Airflow Ollama Jina Langfuse Docker

Replaces manual equity research with a 4-agent LLM pipeline that generates analyst-quality reports on demand.

• 4 specialized LLM agents: Performance Analyst · Market Expert · Report Generator · Critic
• Qdrant semantic caching at 95% similarity threshold — serves cached reports within 24h, cutting redundant LLM calls
• Transfer-learning LSTM: S&P 500 parent model fine-tuned per-ticker with Feast feature store for train/serve consistency
• Full observability: MLflow on DagsHub, Prometheus/Grafana, Evidently AI drift detection, auto-healing on missing models
• Kubernetes manifests + Docker Compose for full reproducibility

Python PyTorch LangGraph MLflow FastAPI Qdrant Docker Evidently AI

• Building Systems with the ChatGPT API — DeepLearning.AI
• LangChain for LLM Application Development — DeepLearning.AI
• Knowledge Graphs for RAG — DeepLearning.AI
• Dataiku Core Designer — Dataiku

B.Tech, Computer Science · Punjabi University, Punjab · 2020–2024 · GPA: 8.97 / 10.0

If you're building something serious in AI — RAG, fine-tuning, agentic systems, MLOps — let's talk.

📧 1001rupindersingh@gmail.com 🔗 linkedin.com/in/rupinder--singh