I'm Jayesh Suryavanshi.
I build LLM agents & GenAI
in production at .

I build predictive models that drive real product decisions, including multi-stage gradient-boosted cascades that refine their predictions as new signals arrive. Trained on distributed data at scale, and treated as products: feature pipelines, monitoring, and evaluation matter as much as the model itself.

Models are only as good as the data behind them. I design and maintain pipelines on Spark, Hadoop, Hive, and Kafka that turn massive, messy datasets into reliable features and signals (batch and real-time) so both ML models and LLM systems have reliable inputs.

I run AI workloads on cloud infrastructure using AWS, Docker, Kubernetes, and Terraform, deploying and scaling services with an emphasis on reliability, observability, and cost. Infrastructure as code by default.

I automate the path to production with Jenkins, Airflow, and Git-based workflows, including CI validation and security scanning for agent configurations, and tooling that generates LLM-readable documentation for codebases. Shipping AI safely should be repeatable, not heroic.

Test coverage across the torch-free core detectors of PyOD (opens in a new tab), the most widely used anomaly detection library in Python: 23 of them are proven to survive a save and load round trip, and 21 a clone round trip. Merged by the library’s creator in August 2026, closing a long-standing issue. PR #708 (opens in a new tab)

GrayEncoder.inverse_transform did not recover the original values it had encoded, so a documented round trip silently returned the wrong answer. Fixed with a regression test and merged by the lead maintainer of scikit-learn-contrib/category_encoders (opens in a new tab). PR #499 (opens in a new tab)

Added a center option to the lPCA intrinsic-dimension estimator, so it can run on uncentered data where the origin carries meaning rather than forcing a recentering that changes the estimate. PR #41 (opens in a new tab)

graphspot (opens in a new tab) is inductive graph anomaly detection with honest flat baselines and a core that does not require torch, which the comparable libraries do without declaring it. It ships flat baselines beside the graph models on purpose, because a detector on raw features is a genuinely hard baseline: on Reddit, the BOND benchmark has LOF outscoring every deep graph model. GeoSocialX (opens in a new tab) maps the geography of geotagged social data and reports its coverage before it maps anything, because most posts carry no location at all. Both install from PyPI.

Twelve further pull requests are open with maintainers, across scikit-learn (opens in a new tab), PyOD, river, imbalanced-learn, hiclass, pythresh, DeepOD and aeon. Four add new detectors, the other eight are correctness fixes. Open is not merged, so I count them separately.