Models served,not notebooksshipped.

The layer between a trained model and a user: feature pipelines, RAG over live data, agentic harnesses, eval loops and serving paths with real latency budgets.

25+

classifiers serving coaching staffs

500ms

P90 on the serving path

2 wks

to earn official Riot data API access

About me

I am Vansh Bordia, an ML engineer focused on the unglamorous middle: the feature pipeline, the retrieval layer, the eval harness, the serving path. A model nobody can query in under half a second is not a product.

My work has shipped as a domain-tuned Gemma layer with RAG over live data, agentic harnesses routing tools and models, and ML pipelines that keep learning as new matches land. Official Riot data API access in two weeks, not two quarters.

Why me

I build the layer between a trained model and a real user: features, retrieval, eval harnesses, serving paths with honest latency budgets, and pipelines that keep improving as new data lands.

I build the layer between a trained model and a real user. A VALORANT-specific Gemma layer with RAG over live patch data, agentic harnesses routing tools and queries, ML pipelines that evolve as matches come in, and eval loops that tell you when a change actually regressed something.

VALORANT-specific Gemma layer with RAG over live patch data

Agentic harness routing tools, models and queries at sub-500ms P90

Type-safe data library on PyPI with retry, rate limiting and validation

Role stack

Python / PyTorch / FastAPI / MLflow / Feature Engineering / Model Serving / Vector Databases / Agentic Systems / LLM Fine-Tuning / RAG Pipelines / Model Evaluation / MLOps / Docker / Kubernetes