~/ramin $ whoami

Ramin Ferdos

Senior AI Engineer

I build production AI systems, from agent orchestration to evaluation.

Senior AI Engineer at Tabiat Makan Industrial Group, an FMCG holding with 130+ brands and 103,000 employees, owning AI and analytics products end to end.

Portrait of Ramin Ferdos

● now: Senior AI Engineer @ TMIG

impact.log
agents.count
14
LLM agents orchestrated in one analytics platform
rag.hit@1
84%
on a hand-labeled 88-question Persian benchmark (95% hit@5)
api.payload
~20×
smaller API payloads after modernizing the ARM dashboards
rank.latency
<50ms
ranking and recommendations over 4.3k–25k SKUs

01 // work

Selected work

Things I've built, latest first. Some are confidential, so they have no public link.

The Watchlight

– Present

Private
  • Python
  • JavaScript
  • PWA

A personal estimation tool that replaces gut-feel project estimates with explainable, uncertainty-aware time ranges (p50/p90), with driver attribution for why a project might slip (novelty, focus fragmentation, external dependencies) and a percentile-based personal calibration engine built on historical actual-vs-predicted ratios.

Key highlights: rebuilt as an offline-first PWA with a hand-ported JavaScript version of the Python estimation engine, a cache-first service worker and localStorage-based calibration, so it works fully offline on iPhone Safari.

Text-to-SQL Eval Harness

–

Public
  • LangGraph
  • Text-to-SQL
  • Evaluation

An evaluation-first study of text-to-SQL agents on the Olist e-commerce warehouse. Built a 100-question benchmark across 5 difficulty tiers, with ground-truth SQL executed against the database and a hand-written result-set scorer (float tolerance, column order, NULLs, ties). Compared four setups: local Qwen2.5-Coder-7B and hosted gpt-4o-mini, each as a single-call baseline and as a LangGraph agent with schema retrieval, capped self-correction, and routing to a specialist join prompt.

Key finding: the hosted routed agent reached 84% execution accuracy vs 79% for the baseline (local: 69% vs 66%; single runs). Auditing my own evaluation found a scorer tolerance bug that marked correct aggregate answers wrong; fixing it and re-grading every run reversed an earlier conclusion that routing hurt the hosted model.

Mosaic HR Directory

–

Public
  • FastAPI
  • PostgreSQL
  • PWA
  • faster-whisper

A mobile-first Persian RTL Progressive Web App for searching and viewing employee profiles across a large industrial holding with ~100,000 employees. Built end-to-end with FastAPI, PostgreSQL, and Vanilla JS PWA — no frontend frameworks. Features JWT auth with refresh token rotation, real-time full-text search, compensation and decree history, and Persian voice search via faster-whisper (CPU-only) with rapidfuzz fuzzy matching.

Key highlights: Persian RTL layout with Vazirmatn font, installable PWA with offline support, service worker update detection, and a data anonymization pipeline for 913k salary records. Demo credentials: Employee code 1 · Password demo1234

Olist Customer Analytics

–

Public
  • XGBoost
  • SHAP
  • MLflow
  • dbt

End-to-end customer analytics on 93,358 customers. Built churn prediction models (XGBoost + Random Forest + SHAP) with MLflow experiment tracking. Designed RFM customer segmentation with actionable business recommendations. Production feature engineering pipeline built with dbt and Prefect.

Key finding: delivery delay is the strongest churn predictor. ROC-AUC improved from 0.53 to 0.767 through iterative feature engineering.

Payment App Analysis

–

Private
  • Python
  • Plotly

This project analyzed *1 (Setare 1), a payment application linked to Hamrah-e-Aval, Iran's main mobile provider, using questionnaire data from both users and the general population. The analysis, conducted in Python, included diverse visualizations such as bar plots, violin plots, heat maps, and interactive maps using the Plotly library. The visualizations were categorized into static and dynamic, with interactive treemaps allowing users to delve deeper into categories based on population size. This comprehensive analysis provided insights into user preferences and behaviors, aiding in identifying areas for improvement. While the data is confidential, the visual assets and plots can be discussed privately.

IAPS Affective Task

–

Private
  • PsychoPy
  • Task Design

In this project I have participated as the developer and implemented the designed task by the Ms. Sharifi in PsychoPy. This task consisted of 36 routines and 3 loops. Each section of this task comprises two subsections: images and statements. Additionally, this task includes two rest periods, one between the image and statement subsections in the training section, and the other between the practice and main testing sections. In each section of this task, instructions are displayed for 2 seconds, and images and statements are displayed for 15 seconds each. During the interpretation stage, the participant has 150 seconds to prove their interpretations.

Affective Priming Task

–

Private
  • PsychoPy
  • Task Design

In this project, I served as the developer, implementing Ms. Sharifi's designed task in PsychoPy. The test comprises 10 routines and 2 loops, divided into two sections: training and main test. During the training phase, participants view 5 word pairs on a white background and identify the valence of the target word using the + (positive) and - (negative) keys. Response time is recorded as the main parameter. The screen displays a black plus sign for 1 second before presenting the words, divided horizontally with one part showing white words on a black background and the other showing black words on a white background.

Research Impact Analysis

–

Public
  • Altmetric
  • CRISP-DM

This Project houses the code and data for a research paper conducted by Ramin Ferdos and Ali Hossein Noorafrooz. In this study, we leverage data from Altmetric to analyze the impact and visibility of research output from Iranian universities on the global stage. Our analysis is based on a variety of metrics, including mentions in news articles, patent creation, policy influence, and more. We have followed the CRISP-DM (Cross-Industry Standard Process for Data Mining) methodology to guide our research.

GANs for Education

–

Public
  • GANs
  • Teaching

In this project, I try to implement and teach multiple and different GANs architectures and tackle well-known challenges in this area. These contents are mostly tightly knitted to the posts shared on the LinkedIn account and can be reviewed as additional resources for those posts.

FX Portfolio Optimizer

–

Private
  • SciPy
  • MetaTrader
  • Optimization

In this project, which is a multiple-stage project, I developed a software for optimizing MPT trading strategy in FX trading. The mathematical optimization which is constrained for this project was done with the SciPy package and resulted in more accurate optimization than Excel's GRG non-linear solver. Also, a comprehensive pipeline for this project was designed to be able to use MetaTrader application outputs. Logging and, data tracking mechanism was also developed and, implemented.

TDD Playground

–

Public
  • Python
  • TDD
  • CI

In this repository, I have implemented the test-driven design pattern for a simple problem within Python. also, a basic test for GitHub workflow to achieve CI is implemented. This project is inspired by the "Learning Test-Driven Development" book by Saleem Siddiqui, all credits are due to this book.

Instagram Like Prediction

–

Public
  • XGBoost
  • EfficientNet B7
  • Web Scraping

Developed a predictive model for Instagram post likes using XGBoost, incorporating image content and performance metrics. Created an Instagram bot for data retrieval, with a scraper targeting high-follower and high-post accounts. Employed EfficientNet B7 for object detection and performed feature engineering to enhance prediction accuracy. Utilized CRISP-DM methodology for data analysis and model tuning, resulting in a robust tool for forecasting post engagement.

Pullup Campaign Analysis

–

Public
  • Machine Learning
  • Marketing Analytics

In this project, I analyzed the pullup campaign with a data science approach. this campaign was held for Zarrin Roya company and operated on Instagram. In this project, I analyzed the data of the campaign, extracted strategic insights which shaped the future marketing plans of the aforementioned company, and trained multiple models for predicting the price of advertising media, influencer, and their situation of profit. Two learning approaches and three approaches and multiple machine learning algorithms were implemented.

PyTorch NN Architectures

–

Public
  • PyTorch
  • MNIST
  • CIFAR

I implemented multiple and famous architectures of neural networks with the PyTorch library. Almost all of the mentioned neural network architectures in skills are implemented in this project. It's worth mentioning that multiple datasets such as MNIST, CIFAR, etc. are used in this project.

Neural Net from Scratch

–

Public
  • PyTorch
  • Neural Networks

I implemented an artificial neural network from scratch with 4 different models, for instances in which I used a deep learning library, I used PyTorch.

#Bekhatereman Campaign

–

Public
  • CRISP-DM
  • Sentiment Analysis
  • Web Scraping

In this project, I analyzed the #Bekhatereman campaign (a CSR Campaign for the Official Distributor of Mitsubishi Motors in Iran) data with the Data Science approach. The goal of this project was to assess the effectiveness of this campaign which was held on Instagram based on advertising cost and sentiment analysis. In this project, I have used the CRISP-DM methodology. Every aspect of this project was implemented from scratch by me the most important section is data gathering which I designed and developed an Instagram bot to retrieve the data necessary for this endeavor. In this project, I have trained more than 250 models with different algorithms and hyperparameters.

02 // experience

Experience

From marketing analytics to leading AI products across a 103,000-employee holding.

industry

  1. Jun 2025 — now

    current

    Senior AI Engineer

    @ Tabiat Makan Industrial Group

    Owning AI and analytics products end to end across a major FMCG holding with 130+ brands and 103,000 employees, from design through production delivery.

    • Architected a 14-agent LLM-powered analytics platform: agent orchestration, routing, guardrails and evaluation.
    • Rebuilt Makan AI's document assistant as a Persian RAG engine over 13 official HR, legal, finance, audit and insurance regulations, using bge-m3 embeddings in SQL Server 2025's native vector type. It reaches 84% hit@1 and 95% hit@5 on a hand-labeled 88-question benchmark.
    • Built PULSE, a behavioral-analytics product for the field sales force, now used by 20–50 managers and featured in the group CEO's holding-wide performance review.
    • Modernized ARM, the group's finance and sales dashboard platform, cutting API payloads ~20× and halving DAX fetch time.
    • Leading Arkan, a financial project-management platform for the construction holding's 9 companies.
    • Earlier: ranking and recommendation pipelines for 4.3k–25k SKUs at sub-50ms latency, the Makan+ HR directory PWA, and a fraud risk-scoring system for the field sales force.
  2. Mar 2024 — Jun 2025

    Data Scientist & Python Developer

    @ Freelance

    • Built predictive and optimization models over high-dimensional financial datasets (200+ features) to support product and marketing decisions for a large-scale payment application.
    • Designed KPI frameworks and analytical pipelines that turn raw transactional data into actionable signals for performance monitoring and forecasting.
    • Developed an AI-driven advertising cost optimization system that integrates real-time signals, improving execution efficiency by ~40% under operational constraints.
  3. Nov 2019 — Mar 2024

    Marketing Data Analyst & Creative Strategist

    @ SISARV Marcom Agency

    Designed, implemented and managed 360° digital marketing campaigns for clients such as Mitsubishi Motors, DS Automobiles, Borgward, Roche and Novo Nordisk. Key member of the #Bekhatereman campaign, winner of the 2019 Golden Hawk Award for Best Digital Campaign.

  4. Jun 2015 — Sep 2015

    System Programmer & Database Designer

    @ Allameh Tabataba'i University

    Designed and implemented a database for the employee health plan, automating its tracking and making it easier for employees to use.

research & mentoring

  1. Aug 2023 — Aug 2025

    Data Science & AI Mentor

    @ Siahkolah Academy

    Gave learners starting out in data science and AI the resources, a structured roadmap and practical industry experience to build their careers, and helped create a collaborative community focused on skill development and international income opportunities.

  2. Oct 2022 — Sep 2024

    Research Assistant

    @ Allameh Tabataba'i University

    Led the indexing initiative for the university's scholarly journals, getting 20+ journals indexed in databases such as Scopus and DOAJ. Streamlined the review process, leaving 6 journals under CSAB evaluation for Scopus inclusion, and was commended by management for exceptional performance.

  3. Aug 2021 — Sep 2023

    Mentor, GANs Specialization

    @ DeepLearning.AI

    Mentored learners and community members in DeepLearning.AI's Generative Adversarial Networks (GANs) Specialization.

03 // stack

Tools & techniques

What I reach for, drawn from the work and projects on this page.

stack.yaml
llm_agents:
LangGraphRAGAgent orchestrationRouting & guardrailsLLM evaluationText-to-SQL
ml_data_science:
XGBoostRandom ForestSHAPMLflowCRISP-DMSentiment analysis
deep_learning:
PyTorchTensorFlowKerasGANsEfficientNetComputer vision
data_platform:
SQL ServerVector search (bge-m3)PostgreSQLdbtPrefectDAX
software:
PythonFastAPIJavaScriptPWATest-driven developmentCI
analytics:
KPI frameworksForecastingOptimization (SciPy)PlotlyWeb scrapingBibliometrics (Altmetric)

04 // education

Education & certifications

degrees

  1. Sep 2018 — Sep 2021

    Master of Information Technology Management

    @ Allameh Tabataba'i University

    gpa 17.71 / 20 (3.684 / 4) overall · 18.56 / 20 in primary courses

    thesis Assessing the Effectiveness of an Advertising Campaign on Instagram Based on Advertising Cost and Sentiment Analysis: Case of CSR Campaign

    talk Lectured at a conference on the gaming industry and its use of AI, ML and data science

  2. Sep 2015 — Feb 2018

    Bachelor of Software Technology Engineering

    @ University of Science and Culture

    project Designed and implemented a Raspberry Pi embedded system that shows useful daily information at the faculty entrance

  3. Feb 2013 — Sep 2015

    Associate of Computer Software Engineering

    @ Shamsipour Technical and Vocational College

    project Designed an information system for a charity organization with .NET and Microsoft SQL Server

certifications

Open one to see its courses and badges.

AI Engineering Professional Certificate IBM Nov 2024
Data Science Professional Certificate IBM Oct 2021
Practical Data Science Amazon Web Services Mar 2021
Generative Adversarial Networks (GANs) DeepLearning.AI Mar 2021
Deep Learning Specialization DeepLearning.AI Mar 2021
Python for Everybody University of Michigan Aug 2020

05 // contact

Let's build something

Have a role, a project or a hard AI problem? Email is the fastest way to reach me.