AI Engineer focused on building production-grade ML systems β from agentic pipelines to deployable APIs.
π Bennett University, India Β· Graduating May 2027
π§ sanskarmodi.dev@gmail.com
π Portfolio Β· LinkedIn Β· Kaggle
Google ADK + ReAct Β· Google Cloud Run Β· FastAPI Β· Snowflake
Two-path hybrid architecture: deterministic fast-path handles ~70% of queries at sub-100ms with zero LLM cost. Agentic path for complex multi-turn reasoning. PII detection/redaction, prompt sanitization, circuit breaker with automatic fallback, 80%+ test coverage across unit + integration + security attack tests.
curl -X POST https://leave-policy-agent-641772618787.us-central1.run.app/chat \
-H "Content-Type: application/json" \
-d '{"message": "What is my leave balance?", "session_id": "demo", "employee_id": "E001"}'FastAPI Β· Qdrant Β· LangChain Β· Next.js Β· Hugging Face Spaces Β· Vercel
Hybrid retrieval: dense embeddings + BM25 + concept co-occurrence graph expansion. Recall@5 = 1.00 on multi-section academic document QA. Citation-aware generation, query rewriting, conversation memory.
β Live App Β· Code
PyTorch Β· MONAI Β· Swin UNETR Β· DVC Β· Streamlit
Multi-modal MRI segmentation on 40GB BraTS 2024 GLI dataset. 99.94% accuracy (avg loss 0.0121). Reproducible DVC pipeline with train/val/test splits. 3D axial/coronal/sagittal visualization.
β Live App Β· Code
PPO Β· PyTorch Β· MLflow Β· DVC Β· Streamlit Β· Yahoo Finance
Bitcoin trading agent achieving 203.8% total return vs 124.1% buy-and-hold baseline. Custom reward shaping with transaction costs and turnover penalty. Automated weekly retraining via single-command DVC pipeline.
β Live App Β· Code
PyTorch Β· MediaPipe Β· OpenCV Β· FastAPI Β· Gradio Β· Hugging Face
Video deepfake detection on FaceForensics++ dataset (1,000 videos/subcategory). 93% accuracy / F1. Visual explainability via heatmaps. Gradio UI on Hugging Face Spaces, FastAPI backend for production use.
sktime β Python's leading time series ML library Β· 50,000+ users Β· 3,000+ GitHub stars
Merged 2 production-ready forecasting algorithms into main branch (PR #7909):
- Bias-adjusted Box-Cox forecaster
- DA-RNN (Dual-stage Attention RNN) forecaster
Full code review, CI pipeline, and type-hint compliance.
- MetaGPT-XL β Production-grade multi-agent system (team of 5, in progress)
- Whitebox Hallucination Reduction in LLMs β Hyperparameter-driven research study (team of 3, in progress)
| Area | Tools |
|---|---|
| LLM & Agents | Google ADK, LangChain, LiteLLM, ReAct Pattern |
| ML / DL | PyTorch, Scikit-learn, MONAI |
| Production | FastAPI, Docker, Google Cloud Run, Azure Portal |
| MLOps | MLflow, DVC |
| UI | Streamlit, Gradio |
| Languages | Python, Java |


