Mahir Malik.

Turning complex data & models into software people actually use.

Machine learning systems, LLM pipelines, and agent-based workflows built around connecting data, models, and software into practical tools.

LLM SystemsAgent WorkflowsProduction AI

Explore work

Available for selected work

Uttar Pradesh / India

About

Building AI systems that retrieve, reason, and act.

I am a ML Engineer

I like figuring out how modern AI systems actually work when you put them together, not just in theory but in real applications. I spend most of my time working across machine learning and AI stacks—especially LLMs, retrieval-augmented generation (RAG), vector databases, and agent-based architectures—where the goal is not just to generate outputs, but to create systems that can retrieve the right context, reason over it, and take meaningful actions.

I approach projects from an engineering perspective: designing end-to-end pipelines that connect data, retrieval, and model behaviour into a coherent system. This includes building LLM pipelines, grounding responses through vector search, and developing agent workflows that can plan, break down tasks, and execute them reliably.

Currently, the direction is toward agentic AI—systems that combine reasoning, memory, and tools to operate with a level of autonomy. Most of this understanding comes from hands-on work, building and refining projects that reflect how these systems are used in real-world scenarios.

Working Principles

Build clearly, so complexity stays understandable.

Ship deliberately, so quality becomes part of the process.

Repeat relentlessly, so every version gets stronger than the last.

Design for clarity, so the interface explains the system.

Keep feedback tight, so iteration turns into real progress.

Build with intent, so every detail earns its place.

Projects

Selected works showcasing engineering depth and problem-solving.

A focused selection of product and machine learning work spanning agent systems, backend orchestration, and causal inference pipelines.

Full archive

AI Infrastructure

Meridian

2026

A split-stack chat routing application built with a Next.js dashboard and FastAPI backend, using LangGraph workflows to manage model selection, provider settings, local chat history, and OpenRouter-powered responses.

Stack

Full-stack

Flow

LangGraph

Storage

Local-first

Project highlights

Next.js dashboard for chat, models, history, and provider controls
FastAPI endpoints for health checks and routed chat requests
LangGraph workflow layer for request orchestration and response shaping

Next.jsFastAPILangGraphOpenRouterTailwind CSSTypeScript

View details Source code

Machine Learning

Treatment Effect Estimation

2026

A causal inference pipeline for IHDP treatment-effect estimation, covering naive baselines, S/T/X-learners, manual DML, EconML estimators, policy analysis, and reproducible evaluation artifacts for ATE and CATE.

Methods

Dataset

IHDP

Outputs

ATE/CATE

Project highlights

Reusable preprocessing pipeline for IHDP with aligned scaling and train/test splits
ATE and CATE estimation with S-Learner, T-Learner, X-Learner, and propensity baselines
Manual DML plus EconML LinearDML and CausalForestDML implementations

PythonEconMLDoWhyscikit-learnSHAPPandas

View details Source code

Blogs

Writing that turns technical shifts into usable mental models.

Writing on model behavior, retrieval systems, evaluation, and deployment tradeoffs behind reliable, production-grade machine learning systems.

Why Efficient Models Are Winning

Why raw parameter count is no longer the main story, and how small, disciplined models are reshaping deployment economics.

Read

Theory to Implementation

Papers, methods, and core ideas rebuilt as readable systems work.

A running set of implementation notes that turns equations and model ideas into code, validation, and practical intuition.

Apr 9, 2026/8 min read/ML Foundations

Linear Regression From Scratch

A first-principles rebuild of linear regression with vectorized gradients, closed-form fitting, standardization, and a direct scikit-learn comparison.

Open

Tech Stack