GenAI / Applied AI Engineer building retrieval, evaluation, and agentic systems in Python on AWS — focused on making AI systems verifiably correct, not just impressive in a demo.

Bengaluru, India · B.Tech CSE (AI & ML), JAIN University — 2026

A cost-aware RAG system benchmarked on MultiHop-RAG (Tang & Yang, COLM 2024). An agentic controller iteratively retrieves across up to 3 hops until evidence is judged sufficient, evaluated against 4 baselines — Dense, Hybrid, Hybrid+Reranker, and a learned cost-optimized Adaptive router — using an independent LLM judge and a sealed, hash-verified holdout split. Scored: Evidence coverage 63.2% → 81.5% (baseline → agentic) on multi-hop questions · Adaptive RAG retains 80% answer quality at 21.8% lower cost 🔗 Live demo Python Qdrant FastAPI Next.js scikit-learn

Statistical release-gating for production LLM prompts. Scores a candidate prompt against production on quality, safety (real AWS Bedrock Guardrails), cost, and latency, runs paired bootstrap significance tests with Holm-Bonferroni correction, and returns a declarative GO / REVIEW / HOLD / INVALID verdict — no manual override. Scored: 5 of 6 real candidate prompts rejected or flagged for review before one passed both a 120-case dev set and a sealed 40-case holdout 🔗 Live demo Python FastAPI Next.js AWS Bedrock Guardrails

Generates synthetic tabular data (Gaussian Copula, CTGAN) and issues an ACCEPT / REGENERATE / REJECT decision across four evaluation pillars — fidelity, rare-case retention, privacy-risk, and downstream ML utility — verified against a cryptographically locked, SHA-256-hashed holdout split. 276 automated tests, 96% coverage. Scored: Gaussian Copula → REGENERATE (recall ratio 0.045 vs. 0.85 required) · CTGAN → ACCEPT (composite score 0.935), on the UCI Adult dataset (48,790 rows) Python pandas scikit-learn SDV Pandera

A classic supervised-learning predictive-maintenance pipeline — Random Forest vs. XGBoost vs. Logistic Regression under class-imbalance-aware training (3.4% failure rate), an F2-tuned decision threshold, 5-fold stratified cross-validation, and live per-prediction SHAP explanations served through FastAPI + Streamlit. Scored: PR-AUC 0.933 · 90.2% recall at a 0.6% false-positive rate on a held-out test set Python scikit-learn XGBoost SHAP FastAPI Docker

A full-stack RAG system answering factual and comparative questions across 6 insurance products straight from PDF brochures — hybrid BGE-M3 dense + BM25 sparse retrieval (RRF), BGE cross-encoder reranking, and Qwen3 32B generation via AWS Bedrock, with a citation-verification step that checks every generated claim against retrieved source text. Scored: Faithfulness 0.936 · Context Recall@4 0.837 · Answer Correctness 0.725 (RAGAS, 30-question benchmark) Python FastAPI Next.js Chroma AWS Bedrock

A LangGraph 4-agent pipeline (Planner → Investigation → Reflection Gate → Resolution) that proposes refund/replacement/escalation decisions — every proposal is validated by deterministic Python policy checks before execution. "LLM proposes, Python validates, Action executes." Scored: 100% issue-classification · 93.94% resolution accuracy · 96.97% final-status accuracy (33-scenario suite) Python LangGraph FastAPI Next.js Docker

An interactive demo exposing training-serving feature skew: a LightGBM fare model scored through duplicated ("Broken") vs. shared ("Correct") feature-transformation pipelines, across 9 features and 3 injectable skew scenarios. Showed: a distance-unit skew silently shifting a predicted fare from ~$20.29 (9/9 parity) to ~$29.19 (8/9 matched) Python LightGBM FastAPI React pytest

A pipeline processing 7.08M U.S. domestic flights (BTS, full-year 2024) through PostgreSQL bulk-loading and tail-number-based aircraft-rotation reconstruction, independently estimating delay propagation and benchmarking it against the BTS's official attribution. Found: 95.28% valid aircraft-link rate · 0.764 buffer-adjusted correlation · downstream delay rate falls from 69.06% to 15.29% as buffer time increases Python PostgreSQL SQL Plotly Dash

• Sales Intelligence Platform — ML lead-scoring classifier (Hot / Warm / Cold) with a Streamlit dashboard.
• Customer Churn Prediction Dashboard — Logistic Regression churn model (ROC-AUC 0.83) with an interactive Streamlit dashboard.

GenAI & LLM: RAG · Multi-hop / agentic RAG · LangGraph · AWS Bedrock (Converse API, Bedrock Mantle, Guardrails) · Strands Agents · Local LLM serving (Ollama) · LLM evaluation (RAGAS, DeepEval, LLM-as-judge) · Statistical release gating (bootstrap CIs, Holm-Bonferroni) Cloud (AWS): Bedrock · EC2 · S3 · SageMaker AI · IAM · DynamoDB ML & MLOps: LightGBM · XGBoost · scikit-learn · SHAP · feature engineering · MLflow · Docker · GitHub Actions (CI/CD) Backend & Data: Python · SQL · FastAPI · Next.js · Pydantic · PostgreSQL · SQLAlchemy · pandas · NumPy · FAISS · Qdrant Visualization: Plotly Dash · Tableau · Power BI