AI Data Scientist @ ACE · MSc AI, VU Amsterdam · Based in the Netherlands

I build LLM systems that actually ship — RAG, knowledge graphs, agentic workflows and MCP — for Dutch financial institutions.
Increasingly interested in the other half of the problem: can we tell what these systems are actually doing - eXplainable, aligned and safe AI

• 🛡️ AI alignment & safety is the problem I care most about — I'm going deep on explainable AI and mechanistic interpretability: SHAP/LIME in production, then attribution, probing, and hallucination detection in LLMs.
• 🧩 Neuro-symbolic AI — my MSc thesis showed that giving an LLM the schema of a knowledge graph before it answers is worth +8.9 accuracy points on EU legal KGQA, across five models (GPT-3.5, Llama-3, Mixtral, …). Structure is cheap; hallucination is expensive.
• ⚙️ Day job: LLM-RAG pipelines, evaluation harnesses, MCP orchestration and knowledge-graph engineering under EU AI Act / GDPR constraints.
• 🎛️ Nights: Philosophy | Psychology | DJing

CLARIAH/CoW — contributor ⭐ 48

An integrated CSV → RDF converter built on the W3C CSVW standard, developed by VU Amsterdam, Utrecht University and the IISH under the NWO-funded CLARIAH project.

I worked on the conversion pipeline during my time at VU Amsterdam, improving CSV→knowledge-graph throughput by 65.3% through memoization and caching, plus an interactive GUI for schema editing.

Languages Python · JavaScript · SQL
AI/ML PyTorch · TensorFlow · HuggingFace · LangChain · Langfuse · W&B · Optuna · SHAP/LIME
Data Neo4j · PostgreSQL · MongoDB · vector DBs · RDF/CSVW
Infra Docker · FastAPI · Azure AI Foundry · Terraform · Pulumi · RabbitMQ · CI/CD
Design Figma · Adobe Suite · Svelte

💬 Ask me about RAG, knowledge graphs, interpretability — or about socrates
📫 pundhirharsh6@gmail.com