I build production GenAI and ML systems — from retrieval pipelines to the cloud-native infrastructure they run on. A lot of my recent work sits at the point where a model's confidence stops matching its correctness, and I like designing the evaluation and tooling that catches that early rather than after deployment.
- 🔭 Currently building GenAI and ML products at Scopic Software
- 🌱 Currently deepening my system design and cloud-native development
- 👯 Open to collaborating on RAG, agents, and applied ML evaluation projects
- 💬 Ask me about GenAI, RAG, and Python
- 📄 More on my background: subash-pandey.com
latent-parking World-model parking planner combining CEM planning, a JEPA latent predictor, and latent MPC — with explicit "distrust gates" that flag when the model's imagined rollout shouldn't be trusted.
cloud-native-AI-platform Cost-bounded AI summarization platform: FastAPI on Kubernetes/Helm, provisioned with Terraform on Hetzner, deployed via GitOps (Argo CD), with a full observability stack.
Activity-Recognition Subject-independent human activity recognition on WISDM. A naive split scores 0.89 macro-F1 — a correct subject-level holdout drops that to 0.29. The project exists to prove that gap and build an evaluation setup that can't hide it.
Gavel A workspace for pressure-testing startup ideas: collecting evidence and running structured multi-agent debate to judge them.
PageAnchor Every answer cites a verifiable region on a page, or it refuses. Hybrid text + visual RAG for layout-heavy PDFs with strict quote verification.
More projects → github.com/notsubash
📫 Email · 💼 LinkedIn · 🌐 Portfolio
⚡ Fun fact: I refactor life decisions.
Core / Web
GenAI / LLM / Agents
Data Science / ML
Cloud & Infra
Databases
Also worked with: Langfuse · Grafana Loki · Tableau · Seaborn · Gephi · Microsoft Office