A production ReAct agent with RAG over a custom PDF parser, a hardened code sandbox, on-demand skills and speech-to-speech voice.

• Built a production AI agent (LangGraph + FastAPI) that plans its own steps, calls tools, writes and runs its own Python, and streams every step back live, with MCP support so it can plug into external tool servers.
• Built the full RAG pipeline: documents are chunked and embedded into Qdrant, then retrieved with hybrid search. Vector search and live web search run at the same time, so one slow or failing source never blocks an answer.
• Wrote a custom PDF parser service to pull clean text, tables and images out of uploaded documents, replacing an unreliable third-party dependency and making document quality something I control instead of hope for.
• Made answers survive a page refresh. The agent now runs as a background job instead of inside the web request, so closing the tab or reloading no longer kills a running answer; you reconnect and pick up mid-stream.
• Built a speech-to-speech voice mode on the same agent: you can talk to it instead of typing, and it talks back, with the same tools, memory and RAG available either way.
• Built AI cost tracking and per-user credit limits on Redis, and instrumented every LLM call with LangSmith for tracing, latency and spend analytics, so cost per answer is measured, not guessed.
• Set up full CI/CD: GitHub Actions runs the test suite on every push, builds a Docker image, deploys to AWS EC2, health-checks it, and rolls back automatically if anything fails.

End-to-end MLOps pipeline with automated CI/CD, a cloud model registry and a deployed inference API.

• Built end-to-end MLOps pipeline predicting customer interest in vehicle insurance using 12 features.
• Implemented automated CI/CD with GitHub Actions, Docker, AWS ECR, and EC2 for model deployment.
• Deployed FastAPI service with S3 model registry, MongoDB Atlas, and schema-driven data validation.

A decoder-only language model built from scratch: modern architecture, pretraining, instruction tuning and preference alignment.

• Built TinyLLM with WordPiece, RoPE, GQA attention, GELU, and pre-LayerNorm in a decoder-only GPT model.
• Trained with mixed-precision next-token LM using AdamW + warmup for efficient single-GPU optimization.
• Added instruction tuning and preference ranking alignment instead of RLHF to refine model behavior.

A risk-calibrated default classifier tuned to the cost asymmetry of missing a defaulter.

• Built a predictive score using XGBoost on 30,000+ records, addressing imbalance via SMOTE.
• Achieved F2-score of 0.603, identifying 84% of defaulters for early warning and loss mitigation.
• Created 8 financial features and tuned threshold aligned with the bank's risk policy.

Deep learning models forecasting firm-level market cap across multi-year horizons.

• Built forecasting models to predict firm-level market capitalization across 1–3 year horizons.
• Developed LSTM, MLP, and Encoder–Decoder architectures on 25 years of historical market data.
• Applied PCA on 28 indicators to reduce dimensionality and improve model generalization.

Fusing satellite imagery with tabular property data to predict real estate prices, with a real ablation study.

• Built a multimodal deep learning pipeline fusing EfficientNet-B0 satellite-image embeddings (1,280-d) with 22 tabular features and learned zip-code embeddings, on ~21,000 properties.
• Engineered KNN-based neighbourhood features (neighbour price, neighbour size) and applied CatBoost for final price regression.
• Reached R² = 0.91 (RMSE ≈ $106K) through systematic ablation, improving from a tabular-only baseline (R² = −18.96) to the full fused model.
• Ranked in the top 2 projects in my college.

Delay severity forecasting over 180K flights, with SHAP-driven operational recommendations.

• Performed EDA on 180K flight records, built predictive models using LightGBM and XGBoost with custom OAI logic.
• Achieved 0.73 ROC-AUC, 0.61 F1-score, and ~30 min MAE for delay severity forecasting.
• Used SHAP for interpretability and proposed strategies to reduce controllable delays across operations.