AI Engineering · Backend Systems

AI systems from model to product.

Turning AI ideas into usable systems: systems that do not only run models, but understand data, use tools, make decisions, and fit into real product workflows.

View projects↘View resume

01 / PROFILE

02 / WORK

Selected projects

Legal RAG reliability experiment

LegalGraphRAG

GitHub ↗

Problem

Legal RAG systems need more than relevant retrieval; they need answers that stay reliable, traceable, and measurable when the legal corpus, citations, and reasoning path become complex.

Solution

A practical research pipeline for testing which interventions most improve Legal RAG reliability, with documented experiments across retrieval, graph structure, reasoning, citation checks, and evaluation metrics.

My contribution

Building the baseline, evaluation workflow, experiment logs, and reliability checks needed to compare retrieval and graph-based methods without changing the base model too early.

Technical challenge

The hardest current challenge is handling heavy evaluation compute. I am working on batching evaluation jobs and moving larger runs to Colab so experiments stay reproducible and affordable.

Baseline architecture

Future architecture

Engineering highlights

Reliability-focused Legal RAG experiments
Planned additions: GNN, NLP, reranking, and USA legal data
Batched evaluation workflow for heavy compute runs

Stack

PythonRAGGraph RAGGNNNLPRerankerUSA Legal DataColab

Status: Reliability experiment in progress

SaaS product engineering

QuestLine

GitHub ↗

Problem

Coaches, athletes, and administrators need one governed workflow for verification, relationships, quests, sessions, attendance, calendar, and payments.

Solution

A role-aware sports SaaS MVP with explicit business rules, multi-role application shells, modular APIs, and a relational product model.

My contribution

Built and documented a TypeScript monorepo across auth, coach verification, athlete relationships, quests, sessions, attendance, calendar, follows, profiles, and the payment MVP flow.

Technical challenge

Keeping authorization, payment activation, attendance eligibility, visibility, and lifecycle transitions consistent across a broad domain model.

Current MVP architecture

Final architecture vision

Engineering highlights

Modular Monolith architecture with clear domain boundaries for Auth, Verification, Relationship, Sessions, Attendance, Payments, and Admin Review
Access/refresh-token authentication, role selection, and role-based access control
Relational Prisma data model for coaches, athletes, coach verification, sports fields, sessions, attendance, payments, calendar, and audit
Persian-first, mobile-oriented, RTL-ready web app with a path toward PWA growth
Separated MVP from future scope such as QP, Wallet, Marketplace, Guild, and ML to avoid premature complexity

Stack

Next.jsNestJSTypeScriptPrismaPostgreSQLTailwindCSSTurborepopnpm workspace

Status: MVP in progress / Operational Core Ready

Human-in-the-loop grading

Teacher-Supervised AI Exam Platform

GitHub ↗

Problem

The full exam lifecycle, from question design to grading, review, result publication, and appeals, needs a system that moves quickly while keeping the final decision with the teacher.

Solution

A class-centered monorepo for exam authoring, scheduling, student-specific access links, deterministic and AI grading, teacher review, result publication, leaderboards, and appeals.

My contribution

Designed and implemented the core domains, modular APIs, Celery queues, grading and AI flows, review system, publishing, appeals, audit logs, and end-to-end tests.

Technical challenge

Separating reliable grading from probabilistic model judgment, managing background jobs, preserving decision traces, and keeping a safe path for teacher review and grade correction.

System architecture

Engineering highlights

Separate queues for AI, grading, email, and leaderboard work
Low-confidence answer review and grade-change history
Mock providers for safe development and tests
Dedicated tokens for exams, results, and leaderboards

Stack

Next.jsFastAPIPythonPostgreSQLRedisCelerySQLAlchemyDocker

Status: Prototype Ready

Multilingual RAG system

Document Assistant

GitHub ↗

Problem

Private Persian and English documents need useful answers without losing page provenance or inventing unsupported claims.

Solution

A Persian-first document intelligence workspace for grounded QA, summaries, comparisons, table questions, multi-turn conversations, and structured AI tools.

My contribution

Implemented the application surface, authoritative request dispatcher, hybrid dense/BM25 retrieval with RRF, bounded reranking, citation validation, document ingestion, streaming, persistence, and evaluation harnesses.

Technical challenge

Routing by intent without blindly executing model output; preserving immutable evidence and physical-page provenance across multilingual retrieval and generation.

System architecture

Document ingestion and indexing

Engineering highlights

Deterministic plan validation and fallback routing
Page-aware PDF/DOCX/TXT ingestion
Gold sets, IR metrics, judge calibration, and regression tests

Stack

PythonFastAPINext.jsPostgreSQLQdrantBM25OpenRouterDocker

Status: Active development

Agentic hiring workflow

Resume Intake Agent

GitHub ↗

Problem

Resume intake spans extraction, duplicate detection, fit screening, approvals, scheduling, and communication across several tools.

Solution

A local LangGraph and LLM pipeline that reads text or scanned resumes, builds a candidate profile, evaluates fit, pauses for Telegram approval, then coordinates the calendar event and invitation email.

My contribution

Designed and implemented an agentic LangGraph pipeline with structured extraction, state-based branching, checkpointed human approval, SQLite memory, and integrations with Google Sheets, Calendar, Gmail, and Telegram.

Technical challenge

Making a multi-service agent resumable and human-controlled while preventing transient failures from losing candidate state.

System architecture

Engineering highlights

Typed Pydantic schemas for extraction and fit assessment
Fallback between text PDF extraction and a vision model
Checkpointed human approval with LangGraph interrupt
SQLite memory for candidate state and duplicate prevention
Gmail sending with fallback to a local draft

Stack

PythonLangGraphOllamaPydanticSQLiteGoogle APIsTelegram

Status: Published

Local agent control plane

Mission Control OS

GitHub ↗

Problem

When a local agent moves into real work, chat alone is not enough; runtime state, memory, services, approvals, conversations, voice, hardware, and multi-step operations need one observable and controllable cockpit.

Solution

A dashboard plugin for Hermes Agent that connects to the runtime through a FastAPI bridge and exposes a live /mission-control panel for snapshots, chat, approval queue, agentic operations, sessions, agent/provider settings, and local voice/TTS.

My contribution

Designed and implemented the plugin architecture, manifest, backend API, dashboard asset pipeline, session-backed bridge, runtime and talk endpoints, operation/event store, approval bridge, Agent/Skill/Tool settings, and local Whisper and Piper hooks.

Technical challenge

The local UI had to integrate with the real Hermes runtime without leaking sensitive state, personal paths, API keys, or approval-gated actions to the frontend or public output; plain chat also had to stay separate from expensive agentic execution.

System architecture

Engineering highlights

Local observability for runtime, vault, services, and hardware
Session-backed bridge for Hermes native conversations
Operation model with event stream and visible pipeline
Approval-aware action flow with the native approval queue
Backend-side API key storage with redacted responses
Configurable hooks for Whisper, Piper, and ffmpeg
Separation between plain chat and agentic operations
Public packaging without hardcoded local paths

Stack

Hermes AgentFastAPIPythonJavaScriptWhisperPiper

Status: Public 0.7, functional prototype in progress

03 / CAPABILITY

Engineering focus

The areas I keep practicing across AI projects, backend systems, and product-shaped prototypes.

Grounded AI systems

RAG pipelines, citation-aware answers, evaluation checks, and clear behavior when reliable evidence is missing.

Controlled model behavior

Placing LLM decisions behind routing, validation, permissions, and product rules instead of trusting raw model output.

Backend product systems

Domain models, APIs, migrations, queues, auth, billing flows, and lifecycle states for real product workflows.

Agent workflows

Stateful orchestration, tool use, human approval points, memory, and recovery from partial failures.

04 / BACKGROUND

About

My path began in psychology. I was looking for a new way to create meaningful work in that field when a machine learning course introduced me to a different direction. The possibilities of AI convinced me to focus on learning one of today's most practical and fast-moving technologies.

Along the way, I have studied Python, statistics, data analysis, machine learning and deep learning, text processing, APIs, databases, Docker, system design, and MLOps. I am especially interested in agentic AI, RAG systems, and practical products built around language models.

I currently turn what I learn into end-to-end practice projects and working prototypes. I approach problems through benchmarks, data-driven tests, and evaluation so that building becomes a way to deepen and verify my understanding, not just produce an output.

I am at the beginning of my professional AI engineering journey and looking for my first opportunity to contribute in a real team, take responsibility, and grow alongside experienced engineers. My psychology background also helps me keep human behavior and real user needs in view while building technology.

Building as practice

Most of my learning happens through building: RAG systems, agent workflows, backend services, and product-shaped prototypes that force ideas to meet real constraints.

Current focus

I am currently deepening my knowledge of agentic AI and RAG by building personal prototypes, testing ideas with measurable criteria, and documenting what I learn.

Build process

I usually start small, make assumptions explicit, measure behavior, study failures, and then harden the system step by step.

Education

B.A. in Psychology, Imam Reza University

05 / TOOLKIT

Technical skills

AI Engineering

LLM applicationsRAGEmbeddingsRerankingAgentic workflowsLLM evaluationGrounding

Backend

PythonFastAPINestJSREST APIsSQLAlchemyPrismaPydanticCelery

Data

PostgreSQLQdrantRedisSQLiteBM25Alembic

Product & Web

Next.jsReactTypeScriptTailwind CSSRTL / i18nSaaS domain modeling

Infrastructure

Docker ComposeGitLinuxLocal LLM servingBackground workers

06 / SOFT SKILLS

Soft skills

The habits I try to bring into teamwork, learning, and real project environments.

Clear communication

I explain decisions, constraints, and blockers early so collaboration stays calm and practical.

Feedback-ready teamwork

I am comfortable asking questions, receiving critique, and turning feedback into concrete next steps.

Fast, disciplined learning

I learn by building small versions, documenting what changes, and testing assumptions against real behavior.

Structured problem solving

I break vague problems into observable parts, compare options, and keep the reasoning traceable.

Ownership

When I take on a task, I care about the outcome, the handoff, and the parts that need follow-through.

User empathy

My psychology background helps me notice where product behavior meets real human needs and confusion.

07 / CONTACT

Let’s talk about AI projects and opportunities to build.

For relevant engineering conversations, reach out by email, LinkedIn, or GitHub.

Email↗LinkedIn↗GitHub profile↗