Portfolio

Mohammad Shojaei

AI Engineer focused on production LLM agents, retrieval systems, evaluation, and applied machine learning.

Email - LinkedIn - GitHub - Hugging Face

Profile

I build production-oriented AI systems: LLM agents, RAG pipelines, OpenAI-compatible APIs, evaluation workflows, and machine-learning tools for real-world domains such as real estate, finance, healthcare, education, and Persian NLP.

My recent work is centered on agentic systems that need to be reliable outside demos: streaming responses, tool calling, conversation memory, retrieval quality checks, structured outputs, tracing, cost and token tracking, regression checks, and deployment runbooks. I work mostly with Python, FastAPI, LangGraph, Hugging Face, vLLM, vector search, reranking, QLoRA/LoRA fine-tuning, and evaluation frameworks.

I care about AI systems that are measurable, debuggable, and honest about their limits. A good system should not only answer; it should expose its sources, failure modes, latency, cost, and quality signals.

Current Focus

Building production LLM agents with tool calling, memory, routing, monitoring, and safe fallback behavior.
Designing RAG and deep-research pipelines with source-grounded answers, reranking, citation filtering, and retrieval evaluation.
Fine-tuning and evaluating Qwen/Gemma-family models for domain-specific behavior, especially Persian and Russian-domain workflows.
Turning prototypes into maintainable services with FastAPI, OpenAI-compatible APIs, streaming, structured logs, health checks, systemd deployment, and benchmark suites.

Experience

AI Engineer, Sunlife Estate

Remote, Moscow - May 2025 - Present

Building a Russian real-estate voice agent that integrates ASR, LLM reasoning, XTTS, and LiveKit telephony for property outreach, qualification, objection handling, callback flows, and specialist handoff.
Developed an OpenAI-compatible backend with streaming responses, tool calling, thread-safe conversation context, token tracking, caching, prompt optimization, detailed monitoring, and persistent memory.
Designed a LangGraph-based conversation engine with intent classification, dynamic routing, busy-user handling, company-information responses, multilingual property-search tools, and DuckDB full-text search.
Fine-tuned and evaluated Qwen/Gemma-family models for real-estate operator conversations, improving brevity, instruction adherence, and sales-script consistency.
Reported sub-250ms p95 latency and about 35% infrastructure cost reduction through streaming, quantization, vLLM serving, model-selection trade-offs, tracing, benchmark suites, and operational runbooks.

AI Engineer, No Limit Markets Ltd

Remote, Dubai - Jan 2025 - Dec 2025

Built AI-powered financial research and market-intelligence agents for stock analysis, financial news monitoring, trading-signal processing, Persian financial content generation, and Telegram-ready reporting.
Developed an 8-stage FinancialKG pipeline for RSS ingestion, web scraping, topic extraction, web search, financial analysis, post generation, Persian translation, and Telegram publishing.
Built a Tehran Stock Exchange multi-agent system with dedicated data, analysis, and content agents for technical analysis, fundamental analysis, risk assessment, and structured financial reporting.
Integrated PyTSETMC, DuckDuckGo Search, Exa Search, RSS feeds, web scraping, Codal/TSETMC-style sources, custom embeddings, and LangGraph tool-calling workflows.
Added retries, checkpointing, link deduplication, admin commands, daily reports, structured logs, error recovery, schema-guided generation, LLM-as-judge review, and report validation.

ML Engineer, Alpha Neuroscience

Tehran - May 2022 - May 2024

Built ML and signal-processing features for ALPHA, a PyQt5 desktop platform for real-time EEG visualization, recording, analysis, quality monitoring, patient management, clinical review, and report generation.
Developed EEG preprocessing workflows with MNE-Python, SciPy, NumPy, and Pandas for EDF loading, filtering, notch filtering, artifact handling, segmentation, and frequency-domain analysis.
Implemented real-time signal-quality and artifact-detection logic covering channel correlation, flatline detection, adaptive thresholds, RMS/PSD checks, eye blinks, EMG, electrode pops, cable movement, baseline drift, and 50/60Hz interference.
Built spectral-analysis modules for FFT, PSD, band-power extraction, Z-scored power ratios, and topographical brain-map visualizations across Delta, Theta, Alpha, Beta, and High-Beta bands.
Built medical assistant and report-support workflows that helped technicians and clinicians summarize EEG findings and review patient context.

Independent AI/LLM Engineer

Remote - 2023 - Present

Built custom RAG, deep-research, and knowledge-management systems using FAISS, Weaviate, LangChain, LlamaIndex, OpenRouter, Ollama, Streamlit, custom chunking, reranking, memory, citations, and persistent Q&A workflows.
Developed document-grounded Persian banking and enterprise assistants with source-aware retrieval, citation filtering, streaming answers, reranking, and strict answer-from-documents guardrails.
Built educational AI systems that convert lecture videos and transcripts into interactive learning materials, including transcript retrieval, summaries, quizzes, and tutorial-style UI generation.
Delivered VLM/OCR and document-extraction pipelines using Qwen-VL, Florence-2, EasyOCR, PaddleOCR, Tesseract, object detection, salary-document extraction, and OCR quality comparison workflows.

Selected Open-Source Work

ReActMCP: MCP web-search server/client that connects Exa-powered search to AI assistants.
pytsetmc-api: typed Python client for Tehran Stock Exchange Market Center data retrieval with Pydantic validation, retries, logging, fallbacks, CLI usage, tests, and documentation.
Hugging Face: 14+ models, Persian NLP datasets, and leaderboard spaces focused on low-resource evaluation, tokenizer training, LoRA adapters, model cards, dataset cards, and Persian benchmark workflows.
Broader GitHub portfolio: 50+ public repositories across LLM tooling, local AI apps, RAG, MCP clients, prompt engineering, web scraping, browser extensions, subtitle translation, VRAM estimation, Persian LLM evaluation, and AI learning resources.

Technical Skills

LLM and agents: LangGraph, CrewAI, AutoGen, tool/function calling, multi-step agents, conversation memory, dynamic routing, intent classification, specialist handoff, ReAct-style workflows, MCP, prompt optimization, guardrails, agent evaluation.

RAG and retrieval: FAISS, Weaviate, LangChain, LlamaIndex, custom chunking, reranking, citation-aware retrieval, source-grounded Q&A, DuckDB FTS, vector search, RAGAS, DeepEval.

Model engineering: Hugging Face Transformers, Qwen, Gemma, QLoRA/LoRA, PEFT, bitsandbytes, Unsloth, W&B, fine-tuning, quantization, model cards, dataset cards, Whisper fine-tuning, TTS.

Production AI and LLMOps: FastAPI, OpenAI-compatible APIs, streaming responses, vLLM serving, token/cost tracking, caching, structured logs, tracing, health checks, systemd deployment, operational runbooks, latency optimization, regression checks, Kubernetes, MLflow, CI/CD for ML.

Data and ML: Python, NumPy, Pandas, SciPy, MNE-Python, signal processing, feature engineering, anomaly detection, FFT, PSD, EEG preprocessing, PySpark.

Automation and integrations: web scraping, RSS ingestion, Telegram bots, MCP clients, SSE/stdio transports, OpenAI function calling, financial-data APIs, OCR/VLM pipelines, structured extraction workflows.

Education and Training

B.Sc. Computer Engineering Higher Education Complex of Bam - 2020 - 2024

LangChain for LLM Application Development, DeepLearning.AI, 2024
CS224N: Natural Language Processing with Deep Learning, Stanford, 2024
Machine Learning Specialization, DeepLearning.AI, 2023
CS50P: Introduction to Programming with Python, Harvard, 2023

What This Site Is About

This site is where I publish practical learning material on LLMs, RAG, agents, evaluation, and production AI engineering. The goal is to make advanced AI engineering easier to reproduce: clear concepts, concrete code, realistic failure modes, and measurable outcomes.