Sayojya Patil

Authenticated Django REST backend for loan submission, document-text extraction, and rule-based underwriting recommendations. Owner-scoped access, duplicate-submission protection, post-commit Celery jobs with transactional claiming and retries, and non-root Docker, covered by 16 tests in GitHub Actions. Historical pre-auth benchmarks: 17,256 requests with zero failures at 297 requests/sec and 21 ms aggregate P95; 97% faster indexed lookups; 100/100 schema-valid optional OpenAI extraction.

• Python
• Django REST Framework
• PostgreSQL
• Redis
• Celery
• Pydantic
• Docker
• GitHub Actions

Uptime monitoring with a FastAPI control plane and Redis-queued workers that run HTTP health checks. Distributed locking kept 100K+ queued jobs free of duplicate runs, and scaling workers from 2 to 8 raised throughput 3.2x to 137 checks/sec at sub-100ms P95 in load tests. Deployed on AWS ECS/Fargate with RDS, ElastiCache, and CloudWatch.

• FastAPI
• PostgreSQL
• Redis
• Docker
• AWS ECS/Fargate
• CloudWatch

FastAPI API that extracts diagnoses, medications, follow-ups, urgency, and confidence from typed clinical notes and uploaded note images with GPT-4o, validating every field with Pydantic and attaching a human-review warning. JWT authentication with bcrypt, account-level data isolation, and validated image uploads, covered by 22 tests in GitHub Actions CI. Built and tested on synthetic data.

• FastAPI
• GPT-4o
• PydanticAI
• SQLAlchemy
• JWT/OAuth2
• pytest
• GitHub Actions

End-to-end analytics app: a Python ETL pages, validates, and upserts 50,000 City of Chicago inspection records into PostgreSQL with retries and transaction rollback. FastAPI serves SQL-aggregate metrics to a React + Recharts dashboard with five KPIs, risk-level failure rates, and ZIP-code trends, covered by 16 tests in GitHub Actions CI.

• Python
• FastAPI
• PostgreSQL
• React
• Recharts
• Vite
• GitHub Actions

FastAPI and PostgreSQL warehouse backend: capacity-limited bins, atomic stock transfers, and an append-only stock-movement audit trail. Velocity-based (A/B/C) slotting, order allocation that prioritizes fast-moving zones, walk-order pick lists, and stale allocations rejected at shipping with HTTP 409.

• Python
• FastAPI
• SQLAlchemy 2.0
• PostgreSQL
• Docker Compose
• pytest

Security study of a RAG pipeline (MiniLM embeddings, FAISS top-5 retrieval, FLAN-T5) across 50 questions and a 500-passage corpus. Corpus poisoning cut answer accuracy from 88% to 58% (26% attack success) versus 4% for prompt injection, though both reached 100% retrieval recall. A keyword filter blocked the tested injections, while a retrieval-score anomaly detector made poisoning worse (26% to 34%).

• Python
• Sentence-Transformers
• FAISS
• flan-t5

Comparative study of seven sequence models (RNN, LSTM, GRU, Transformer, CNN-BiLSTM-Attention) against random-walk and moving-average baselines on historical stock data. LSTM reached 78% directional accuracy versus 51% for the random walk, GRU achieved the lowest error (R² 0.997), and gated recurrent models consistently outperformed the Transformer.

• Python
• TensorFlow
• Keras
• scikit-learn
• pandas

Team big-data project analyzing U.S. domestic flight delays with Apache Spark. My part covered data collection, the temporal-variability analysis, and the visualizations, which showed higher delay exposure for afternoon and evening departures; the team also compared airlines and airports and tested Spark MLlib regression models.

• Python
• PySpark
• Spark MLlib
• pandas
• Matplotlib