NikunjAgrawal

May 2025 — Aug 2025

Provost's Office

University of Pennsylvania

ML Research Engineer

• Built responsive React dashboards and versioned TypeScript REST APIs with Zod validation, rate limiting, and OpenAPI documentation, surfacing ML enrollment insights across Penn's 12 academic schools for 3 downstream teams.
• Fine-tuned a BERT-based classifier on 50K labeled enrollment records with HuggingFace Transformers, reaching 84% held-out accuracy with cross-validation and hyperparameter tuning.
• Stood up an MLflow pipeline with a model registry, retraining triggers, and a versioned serving API; rewrote batch inference with asyncio concurrency for a 6x throughput gain at identical accuracy.
• Added SHAP feature importance, early stopping, and calibrated probability outputs to every deployed model, and set up GitHub Actions CI enforcing type-safety, lint, and test coverage on every pull request.

Jan 2023 — Feb 2024

Technology and government consulting

GDi Partners

Data & ML Consultant

• Built a full XGBoost pipeline with feature engineering, k-fold cross-validation, SHAP explainability, and calibrated REST API serving for a $400M Gates Foundation education program.
• Model-driven interventions informed planning across 900+ government schools, improving learning outcomes by 28% and reducing dropout by 15%.
• Handled severe class imbalance in rare-event prediction with BorderlineSMOTE and threshold tuning across heterogeneous state datasets, and designed asyncio-parallelised feature pipelines over 50+ sources with schema validation and dead-letter queues.
• Built a FastAPI backend with MERN and AngularJS frontend modules for a civic-tech grievance platform with role-based access control, serving 150K+ users.
• Shipped a reusable TypeScript component library with Storybook documentation, cutting new dashboard build time for the team by 60%.

Dec 2021 — Jan 2023

HR SaaS unicorn

Darwinbox

AI Product Engineer, Founder's Office

• Architected a production RAG support chatbot over enterprise knowledge bases with a GraphQL data layer, LangChain backend, and FAISS retrieval, cutting average resolution time by 40% and lifting adoption 25%.
• Fine-tuned a DistilBERT intent classifier on 200K proprietary HR queries with custom tokenization and domain vocabulary expansion, improving intent accuracy 32% over the base model across 10K+ daily queries at sub-2s p99.
• Built an LLM evaluation framework with RAGAS, BERTScore, A/B harnesses, and regression suites, then shipped the inference endpoints it gated.
• Built an ML monitoring stack with drift detection, latency alerting, and weekly accuracy regression reports, plus a WebSocket event streaming layer with TypeScript observability dashboards tracking usage and SLA metrics across 50+ tenants.

Jun 2021 — Dec 2021

Credit fund, London

Astra Asset Management UK

Technical Product Analyst

• Designed a Bayesian multi-objective optimization pipeline in PyTorch and BoTorch using Gaussian Process surrogates and custom acquisition functions for portfolio risk parameter calibration across correlated asset classes.
• Built React and FastAPI tooling exposing optimization results through interactive charts with real-time data refresh.
• Containerized the model-serving stack with Docker and set up AWS CI/CD with lint, test, and staged deploy gates, cutting release cycles to 15 minutes, with automated retraining triggers for production refreshes.