• I build production-grade AI systems that combine Machine Learning, Generative AI, and scalable infrastructure to solve high-impact, real-world problems.
• With 8+ years of experience, my work has evolved from classical ML systems to end-to-end Agentic AI architectures, where LLMs are orchestrated with tools, memory, and workflows to automate complex decision-making.
• Currently, I work as a Lead Data Scientist, designing systems that power credit risk, fraud detection, and financial intelligence at scale, supporting 5M+ predictions per month across an NPR 9B+ lending portfolio.
• I am also actively developing SystemX, a Linux server management software, and MLfy, an Auto ML platform to streamline machine learning workflows.

Designing multi-step reasoning systems using:

• LangChain, LangGraph, CrewAI, Google ADK, MCP

Building RAG pipelines:

• Hybrid retrieval (BM25 + dense embeddings)
• Vector stores (FAISS, Chroma, Elasticsearch)
• Re-ranking strategies for improved grounding

Implementing:

• Tool calling agents (APIs, DBs, credit bureau parsing)
• Memory (short-term + long-term + semantic retrieval)
• Structured outputs (Pydantic, JSON schema validation)

Use cases:

• Automated credit report analysis
• Risk summarization agents
• Decision-support systems

End-to-end ML lifecycle:

• Feature engineering → model training → evaluation → deployment → monitoring

Models:

• Gradient Boosting: XGBoost, LightGBM, CatBoost
• Statistical: GLM, survival models, Bayesian models (PyMC)
• Deep Learning: LSTM, Transformers (NLP tasks)

Core applications:

• Credit risk modeling (PD, LGD proxies)
• Fraud & anomaly detection (PyOD, isolation forests)
• Recommendation systems
• Time series forecasting (Prophet, LSTM)

Metrics:

• AUC, KS, Gini, Precision-Recall, calibration

Pipeline orchestration & tracking:

• Airflow, Kubeflow Pipelines
• MLflow, Git, DVC

Deployment & Scaling:

• FastAPI, Flask (REST APIs) for batch + real-time inference systems
• Docker, Kubernetes

Monitoring & CI/CD:

• Data drift, concept drift, performance monitoring
• Automated training + deployment workflows

• Distributed processing: Spark, PySpark, Hadoop ecosystem
• Streaming: Kafka, Flink
• Storage & querying: PostgreSQL, MySQL, ClickHouse, Cassandra
• Data lake: S3, MinIO, Iceberg
• ETL/ELT: dbt, Airflow pipelines
• Event-driven architectures: Real-time fraud detection pipelines

flowchart TD
    A["User / API Request"] --> B["API Layer (FastAPI)"]
    B --> C["Orchestrator (LangGraph / CrewAI)"]
    C --> D["LLM (OpenAI / Google GenAI)"]
    C --> E["Tools Layer"]
    E --> F["Vector DB (FAISS / Chroma)"]
    E --> G["Feature Store (Feast)"]
    E --> H["External APIs / Credit Bureau"]
    C --> I["Memory Layer"]
    I --> F
    D --> J["Structured Output"]
    J --> K["Decision Engine"]
    K --> L["Database / Downstream System"]

• End-to-end ML system design (not just notebooks)
• Agentic AI workflows with real-world use cases
• Scalable RAG pipelines with evaluation
• Production-ready FastAPI services
• Clean, modular, and reproducible code

"If it’s not in production, it’s a prototype."

• Focus on scalability, reliability, and measurable impact
• Prefer simple, explainable systems over unnecessary complexity
• Treat data + models + infra as one system

If you're working on Agentic AI, RAG systems, or Scalable ML infrastructure, I’m always open to collaborating or exchanging ideas. Let's connect!