Architecting Production-Grade AI Systems

Machine Learning Engineer specializing in building scalable, production-grade AI systems from foundational mathematics to live deployment. Architecting deterministic data pipelines, compound RAG frameworks, and autonomous multi-agent systems designed for complex enterprise environments.

Ecosystem Core Nodes //

NODE // HACKATHONS

AzkiVam FinTech Hackathon

The AzkiVam FinTech Hackathon (December 2025) was a prominent industry event focused on developing innovative and intelligent solutions within the financial eco...

NODE // KAGGLE

Natural Language Processing with Disaster Tweets

This notebook implements a high-scoring text classification model for a competitive Kaggle challenge. It utilizes a specific compressed, lightweight variant of ...

NODE // KAGGLE

Cooking with RAG — A Retrieval-Augmented Recipe Assistant

This notebook builds an intelligent, conversational culinary companion using Retrieval-Augmented Generation (RAG). By coupling a dense vector database of divers...

Systems Ecosystem

Interactive registry of deployed systems, machine learning models, and validation benchmarks.

NODE // OFFLINE // REGISTRY

Simia-MatchMaking

NODE // AGENT

The AI-Driven Therapist Matchmaking Engine is designed to solve the problem of connecting patients seeking mental health care with the most clinically compatible, logistically available, and relationally aligned therapist.

This system resolves these inefficiencies by executing an automated, multi-layered pipeline ( 6 layers). It instantly translates unstructured patient narratives into structured categorical profiles, runs lightning-fast deterministic exclusion queries, and ranks providers using a hybrid approach of explicit algorithmic clinical scoring and optimized classical machine learning for interpersonal alignment.

#RecommenderSystem

View Project↗

NODE // OFFLINE // REGISTRY

Tehran Apartment Price Prediction

NODE // GITHUB

A data-driven machine learning project for estimating the value of apartments located in Tehran, Iran.

All necessary data was collected from Divar — a major Iranian classified ads platform — using a custom web scraper. The dataset underwent thorough cleaning and preprocessing before being used to train regression models that predict apartment prices based on features such as area, number of rooms, floor, and neighborhood.

The project covers the full ML pipeline: data collection via Divar API, data cleaning, exploratory data analysis (EDA), feature engineering, model training, and evaluation.

#Python#Machine Learning#Scikit-learn#Pandas#EDA#Real Estate#Data Science

Source Code→

NODE // OFFLINE // REGISTRY

Multi-Agent Travel Planner

NODE // AGENT

A practical multi-agent system that helps travelers choose when to travel, where to stay, and which flight to book by evaluating complete itineraries instead of optimizing each part in isolation.

Rather than picking the cheapest flight or the best hotel separately, the system scores full trip plans using weather conditions, prices, availability, and personal preferences — then recommends the most reasonable options with clear explanations.

The request flows through a set of specialized agents:

• TravelProfileAssistant — understands traveler habits, budget, and preferences

• WeatherAnalysisAgent — evaluates which dates have better conditions

• FlightOptimizationAgent — finds efficient and affordable flights

• HotelRecommendationAgent — selects suitable accommodations

• Aggregator (Decision Engine) — combines all signals, builds full trip options, scores and ranks the best matches

Instead of guessing, the system compares trade-offs and explains why a certain plan is better.

#Python#Multi-Agent#LLM#FastAPI#Travel#AI

Source Code→

AzkiVam FinTech Hackathon

Synapse//December 2025

NODE // HACKATHONS

The AzkiVam FinTech Hackathon (December 2025) was a prominent industry event focused on developing innovative and intelligent solutions within the financial ecosystem.

The intensive competition aimed to optimize LendTech services and leverage emerging technologies to resolve complex market challenges.

The submitted project presented an innovative decentralized finance (DeFi) solution specifically tailored for crowdfunding.

The architecture established a transparent and efficient framework for smart capital acquisition and management, supported by a functional Minimum Viable Product (MVP).

Following rigorous evaluation among numerous specialized teams, the project successfully secured third place in the competition.

#FinTech#DeFi#crowdfunding

Verify Resolution↗

NODE // OFFLINE // REGISTRY

Cooking with RAG — A Retrieval-Augmented Recipe Assistant

NODE // KAGGLE

This notebook builds an intelligent, conversational culinary companion using Retrieval-Augmented Generation (RAG). By coupling a dense vector database of diverse cooking recipes with a Large Language Model (LLM), the project demonstrates how to overcome standard model hallucinations and knowledge cutoff limitations. It implements efficient semantic search to query recipes by title, preparation method, or "what's in my fridge" ingredients, providing accurate, highly relevant recipes and contextual substitution advice in real-time.

#RAG#LLM#FAISS#semantic-search#Python#Kaggle

Kaggle Project↗

NODE // OFFLINE // REGISTRY

Natural Language Processing with Disaster Tweets

NODE // KAGGLE

This notebook implements a high-scoring text classification model for a competitive Kaggle challenge. It utilizes a specific compressed, lightweight variant of the BERT architecture (bert_en_uncased_L-10_H-128_A-2) with 10 layers, a hidden size of 128, and 2 attention heads. The project focuses on fine-tuning this compact transformer model to achieve strong competitive performance ( Public Score 0.79711 ) while keeping resource consumption and training time minimal.

#NLP#Transformer#BERT#PyTorch#text-classification#Kaggle

Kaggle Project↗

View Full Systems Registry →

Research & Validations

PUBLISHED IN // PUBLICATIONS

CivilCFRP

Experimental study on the behavior of thermally insulated RC beams strengthened with CFRP after their exposure to high temperatures compared to non-insulated ones

// CORE OBJECTIVE:With regard to the expansion of the use of carbon fiber reinforced polymer (CFRP) in strengthening civil engineering structures due to its high positive points (like high tensile strength and low thickness) as well as its weaknesses in high temperatures especially in buildings and weak points of existing thermal insulators, the experiments mentioned in this article have been carried out to investigate the post-fire conditions of CFRP retrofitting systems using the externally bonded reinforcement technique which resulted from the need to use insulation for this type of reinforcement system to improve its heat performance, as well as the weak points of common insulations. In the first phase, 12 samples of reinforced concrete (RC) beams strengthened with externally bonded carbon fiber reinforced polymer (UD200) were heated at 400 °C, 500 °C, 600 °C, and 800 °C and loaded after cooling, then they were compared with the results of the second phase of the tests which have been explained completely, consisting of 11 RC beams strengthened with CFRP having exactly the same properties as those in the first phase. They were also thermally insulated with intumescent paint that had some advantages like low thickness (1.1 mm) and the speed and ease of implementation and restoration. These results have clearly shown that the new insulating layer not only can maintain the positive feature of common insulations, but also unlike other common insulators, does not add to the thickness of the specimens. Moreover, the application of the intumescent paint both increased the performance of the specimens at high temperatures and covered the weaknesses of CFRP reinforcement system against heat so that the CFRP sheets unlike the ones on the non-insulated specimens did not completely disappear at the highest temperature.

Access Source ↗

VALIDATIONS // REGISTRY

Neural Networks and Deep LearningDeepLearning.AI/ Coursera

Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and OptimizationDeepLearning.AI/ Coursera

Natural Language Processing with Attention ModelsDeepLearning.AI/ Coursera

Show More Validations →