I build across the stack — ML systems, backend infrastructure, data pipelines, and real-time applications. Some projects are deeply statistical (causal inference, drift detection, experiment design), others are systems heavy (distributed job queues, neural retrieval, training infrastructure). What ties them together is a bias toward production: things that deploy, scale, and come with real evaluation metrics.

ML / AI

Languages

Backend & Infra

Tools

Automated benchmarking and evaluation framework for RAG pipelines

RAGScope provides a reproducible, quantitative harness for comparing RAG configurations across chunking strategies, embedding models, retriever architectures, and LLMs before any pipeline reaches production. Evaluation runs are fully parameterized and logged against standard metrics — BLEU, ROUGE, context recall, faithfulness — enabling controlled ablation across pipeline components. Built on LangChain with support for multiple vector DB backends, it treats RAG evaluation as a first-class engineering concern rather than an afterthought.

Python LangChain FAISS Vector DBs LLM Evaluation BLEU ROUGE HuggingFace

Two-stage neural retrieval system with LambdaRank reranking on MS MARCO

Atlas Search implements a full retrieval stack: dense first-stage retrieval using FAISS with a trained SentenceTransformer bi-encoder, followed by feature-based Learning-to-Rank reranking via LightGBM LambdaRank, with an optional cross-encoder and stacked reranking stage. The system includes full MRR@10 / nDCG evaluation against the MS MARCO passage ranking benchmark, per-stage inference latency benchmarking, and a FastAPI serving layer. Retrieval quality and latency tradeoffs are measurable at each stage of the pipeline.

Python FAISS SentenceTransformers LightGBM LambdaRank Neural Retrieval FastAPI MS MARCO

Production-grade A/B testing and causal inference engine

ExperimentIQ handles the full experiment lifecycle from sequential testing with O'Brien-Fleming alpha spending bounds to CUPED variance reduction for improved statistical power on high-noise metrics. Beyond frequentist A/B, it exposes causal forest estimation via EconML, Double ML for heterogeneous treatment effects, and Synthetic Control for time-series quasi-experiments. A FastAPI backend surfaces experiment configuration and results, with a Streamlit dashboard backed by Plotly for live monitoring of p-values, uplift estimates, and stopping criteria.

Python EconML Causal Forest Double ML CUPED Sequential Testing FastAPI Streamlit

Production ML system for automated drift detection and self-healing model retraining

DriftGuard-ML implements a fully automated validate → drift-check → retrain → commit → serve loop orchestrated entirely on GitHub Actions, with zero paid services. Drift is quantified using PSI and KS statistics against reference distributions; data quality gates run through Great Expectations before any retraining is triggered. sklearn and XGBoost pipelines are retrained on drift detection, versioned as git artifacts, and hot-swapped into a FastAPI inference service — turning model maintenance from a manual process into a monitored, auditable pipeline.

Python PSI KS Test Great Expectations XGBoost FastAPI GitHub Actions MLOps

Production-oriented ML training infrastructure with distributed training and experiment tracking

ScaleTrain provides a clean, opinionated training framework that scales from single-process to multi-GPU distributed training via PyTorch DDP without changing the training contract. MLflow is integrated at the run level for metric logging, hyperparameter capture, and model artifact versioning, enabling reproducible experiment comparison across hardware configurations. A FastAPI inference service and Docker support complete the path from training run to served model, with strict separation between training, data, tracking, and inference concerns.

Python PyTorch PyTorch DDP Distributed Training MLflow FastAPI Docker

Self-hosted background job processing system with DAG orchestration and real-time observability

Taskfire is a full-stack background job system: jobs are submitted via a typed REST API, persisted in PostgreSQL, and dispatched through a Redis sorted-set priority queue to a pool of dynamically-scaling Go goroutines. DAG-based dependency enforcement ensures correct execution ordering, with exponential backoff retry semantics on failure and a dead-letter queue for terminal failures. A React 18 WebSocket dashboard provides live visibility into queue depth, worker utilization, and throughput; Prometheus scrapes /metrics on a 10-second interval with 15-day TSDB retention, Grafana-ready out of the box.

Go PostgreSQL Redis DAG Goroutines Prometheus Grafana React 18 WebSockets

Real-time collaborative code editor with conflict-free multiplayer editing

CodeCollab delivers a live, multi-user code editing experience with real-time cursor synchronization, conflict-free collaborative editing, and syntax highlighting over a WebSocket-based multiplayer architecture. The system handles concurrent edits without locking, keeping editor state consistent across all connected clients in real time.

TypeScript WebSockets Real-time Collaborative Editing Syntax Highlighting