ram sankar

Built a production computer-vision pipeline for utility-pole inspection, detecting conductor / wire polylines from drone and ground-level imagery on the EPRI Distribution Inspection dataset.

Evaluated and benchmarked segmentation and line-detection models including YOLOv8-seg, SAM 2, and HAWP to scope an aerial-vs-ground imagery detection strategy.

Engineered a Mapillary street-view scraper in Python to source and geo-tag pole imagery, automating dataset collection for model training.

Developed PyTorch training and inference workflows with Docker-based reproducible environments for distribution-grid asset analysis.

Architected a production GraphRAG pipeline using LangChain, Neo4j, and OpenAI embeddings to power a multilingual chatbot serving 50K+ monthly queries with relationship-aware context retrieval.

Designed an end-to-end RAG flow with document chunking, embedding indexing, vector similarity matching, and Cypher graph traversals, improving answer accuracy by 28% over flat RAG baselines.

Built a FastAPI inference backend with async endpoints, Pydantic validation, and streaming responses for low-latency LLM serving in production.

Engineered context-preserving prompt templates with few-shot examples, output guardrails, and structured output parsing for grounded, hallucination-resistant answers.

Implemented a multilingual NLP pipeline with language detection and translation APIs, extending accessibility to Spanish-speaking communities.

Deployed the AI stack on AWS EC2 and Azure App Services with Docker, model versioning, and A/B testing infrastructure.

Built NLP intent classification and context assembly modules for the IRIS AI classroom assistant, processing natural language student queries against a curated knowledge base across 50+ school deployments.

Integrated AI-powered search and recommendation features into a contract management SPA, leveraging API-driven inference endpoints to surface relevant clauses and reducing manual review time by 25%.

Designed data preprocessing and evaluation pipelines for client-facing AI features with structured output validation, confidence scoring, and error recovery patterns.

Delivered real-time AI responses via WebSocket/MQTT integration for interactive learning experiences across physical and digital touchpoints.