in the name of Allah sma-jafari

AI agents.
From code to production.

I build the AI agent layer for real companies.
Software, deployment, and infrastructure.

Explore the engineering

I’m currently responsible for building AI agents at Pricepoint.

Mohammad Ali (sma-jafari)

In the name of Allah. Hi everyone, I’m Mohammad Ali. I build AI agent systems for companies, from the agent software to deployment and infrastructure.

For me, the interesting part is making the whole system work together: giving an agent the right tools, passing information between steps, and handling the moments when something fails or needs clarification. I care about those details, from the first line of code to a running service.

My background is in computer engineering, and my PhD research focuses on AI for startup valuation. Along the way, I’ve built multi-agent frameworks and voice agents. I’m currently responsible for building AI agents at Pricepoint.

This website is a look at how I think and what I build. Feel free to explore my work, read about my background, or connect with me. Thanks for being here.

Conceptual system All layers Task → agent → tools → result. Task Runtime Release

AI agents.
Every layer.

From the agent software to the infrastructure it runs on.

Whole system Agent software A task uses a tool, then returns a checked result. Task Runtime Release

What the agent does.

I build agent workflows with tools and shared state, from Emerald Core to voice agents for phone calls.

Give each workflow its own context.
I select knowledge modules by workflow, then check model output against a schema and application rules.
Keep approvals current.
Approval targets an exact version. Changed inputs mark dependent artifacts stale, requiring review.
Make retries mean one request.
Exact retries return the existing run. The same request ID with different input is rejected.

Current work

I’m currently building AI agents at Pricepoint, a revenue management system (RMS) provider for hotels.

More about Emerald and IDEA SPARKS

From code to a running service.

I package the API, worker, and dependencies together, with explicit checks before startup.

One image, separate processes.
My multi-stage Docker build packages the frontend and Python service. API and worker share one image and a non-root runtime.
Budget retries before startup.
Startup requires worker ownership to outlast all configured model timeouts and retries, plus a safety margin.
Check the dependency, too.
Liveness checks the API process. Readiness checks the database and returns HTTP 503 if it fails.

What I test

I test duplicate submissions, expired worker claims, and progress reconnection alongside backend, frontend, and browser checks.

Where the agent runs.

I build the server side around the agent: APIs, background workers, persistent run state, and operational logs.

Recover with attempt fencing.
Recovery assigns a new attempt number. Writes from the abandoned attempt are rejected, blocking stale run updates.
Commit completion together.
Final run state and its completion event share one Firestore transaction: both commit, or neither does.
Resume from the last event.
Persisted events resume from Last-Event-ID or a saved cursor. Reconnection replays progress without resubmitting the task.

Trace without exposing content

Request and run IDs link structured logs. Prompts, credentials, and private workspace content stay out.

How I package the service

Three connected parts of the agent systems I build.

Mohammad Ali

Known online as sma-jafari.

I build AI agent systems for companies, from agent software to deployment and infrastructure.

My background is in computer engineering. My PhD research focused on AI for startup valuation.

Connect on LinkedIn
Mohammad Ali (sma-jafari)

Experience

Near-Dynamics

Built an AI agent framework from the ground up and developed a voice agent that receives phone calls and handles complex caller requests.

Lab for AI Applications in Business Studies

Created a cognitive scoring model that evaluates candidates through their performance in cognitive games, helping HR teams assess specific skill dimensions.

Suaka

Developed embedded AI for a security system on an STM32 microcontroller. Implemented human-versus-animal detection in C++ from scratch under constrained resources, with alerts when a person crossed the sensors.

Selected projects

Emerald

Multi-agent framework

Created Emerald Core, a custom framework for building multi-agent AI systems. It decomposes complex problems into specialized agents with defined tools, roles, and interaction protocols.

Digital Transformation Assistant

Built with Emerald to support decisions using global strategies and current macro KPIs.

IDEA SPARKS

Built a multi-agent application based on Emerald to help accelerators and startup studios validate startup ideas before launch. Macro, Meso, and Micro agents evaluate different aspects of an idea.

Education

Computer engineering

BE · University of Tehran

My studies at the University of Tehran also include an MA.

Research on AI
and evaluation.

My PhD research focuses on AI for startup valuation. My work also examines research workflows and labor-market data.

Published as

Seyed Mohammad Ali Jafari

ORCID ResearchGate

Publications & preprints

Newest first

Working paper

Founders or ventures? Horse signals dominate jockey signals in early-stage startup screening

Seyed Mohammad Ali Jafari, Ali Mobini Dehkordi, Ehsan Chitsaz, Yadollah Yaghoobzadeh

SSRN

Compares founder and venture signals in early-stage startup screening using 7,958 technology startups and data on funding, teams, media, patents and narratives.

SSRN DOI
Preprint

Deconstructing the Crystal Ball: From Ad-Hoc Prediction to Principled Startup Evaluation with the SAISE Framework

Seyed Mohammad Ali Jafari, Ali Mobini Dehkordi, Ehsan Chitsaz, Yadollah Yaghoobzadeh

arXiv:2508.05491

Reviews 57 empirical studies and proposes a five-stage framework connecting startup evaluation goals, data, features, validation and risk-aware interpretation.

arXiv PDF
Preprint

What Matters Most? A Quantitative Meta-Analysis of AI-Based Predictors for Startup Success

Seyed Mohammad Ali Jafari, Ali Mobini Dehkordi, Ehsan Chitsaz, Yadollah Yaghoobzadeh

arXiv:2507.09675

Synthesizes 58 predictors across 13 studies using a Weighted Importance Score, then examines how predictor rankings change with the definition of startup success.

arXiv PDF
Conference paper

Nasdaq-100 Companies’ Hiring Insights: A Topic-Based Classification Approach to the Labor Market

Seyed Mohammad Ali Jafari, Ehsan Chitsaz

Springer · ODSIE 2023 · CCIS 2205, pp. 98–113

Uses structural topic modeling to organize Nasdaq-100 companies’ online job vacancies into 13 categories and examine patterns in labor-market demand.

Publisher arXiv PDF
Journal article

Artificial Intelligence (AI) vs. Human-Led Approaches in Human Resource Recruitment Assessment: A Meta-Synthesis of Advantages and Disadvantages

Mohammad Etemadi, Ehsan Chitsaz, Sahar Koushki, Seyed Mohammad Ali Jafari

Journal of Sustainable Human Resource Management · 6(11), pp. 191–214 · In Persian

Synthesizes the trade-offs between AI-based and human-led recruitment assessment, including decision quality, acceptance, bias and information security.

English title shown; the article is published in Persian.

Journal record DOI Author copy
Preprint

Streamlining the Selection Phase of Systematic Literature Reviews (SLRs) Using AI-Enabled GPT-4 Assistant API

Seyed Mohammad Ali Jafari

arXiv:2402.18582

Presents a GPT-4-assisted literature-screening workflow covering data preparation, article assessment and structured results.

arXiv PDF
Journal article

Turning the Risks Leading to Failure into Success in Startups (Case study: Startups Located in Tehran and Qom Universities)

Seyed Mohammad Ali Jafari, Zahra Aghamir Mohammad Ali

Journal of Economics and Administrative Sciences · 5, Supplement 2

Examines startup failure risks through interviews with 15 startups, combining competitive pressures with organizational and management factors.

Author copy