8+ years building production-grade AI systems | Tehran, Iran
Specializing in Agentic AI, LLMs, RAG systems, evidence-first automation, and production ML pipelines
I architect and deploy production-grade intelligent systems that drive real business impact. My focus areas:
| Domain | Expertise |
|---|---|
| Agentic AI | Autonomous agents, multi-step reasoning, task orchestration, tool integration |
| LLMs & NLP | Fine-tuning, RAG architectures, NL2SQL, conversational AI, prompt engineering |
| Evidence-Driven AI | Audit trails, evaluation gates, repair proposals, human-in-the-loop workflows |
| Enterprise RAG | Hybrid retrieval, re-ranking pipelines, knowledge management systems |
| Production ML | End-to-end pipelines, real-time inference, MLOps, scalable deployments |
🤖 Production-Grade Agentic AI Framework
Vision + LLM + Event Sourcing • Local LLMs • LangGraph • HITL Safety • Autonomous Task Execution
ARIA is not a prompt-chain demo or a single-purpose script — it's a full agentic AI system built for real-world automation: observe UIs with vision, plan with LLMs, act safely with human oversight, and learn from outcomes. Designed to run on local LLMs and consumer GPUs (8GB VRAM), with native English & Persian support for privacy-sensitive and resource-constrained environments.
Cognitive architecture — perception, reasoning, execution, and memory are separated and observable:
┌─────────────────────────────────────────────────────────────────────────┐
│ ARIA — Cognitive Core │
├─────────────────────────────────────────────────────────────────────────┤
│ 👁️ Eye (VLM/OCR) → Observe real interfaces • Screenshot • UIRef │
│ 🧠 Brain (LLM) → Plan, execute, observe • LangGraph • HITL gates │
│ ✋ Hand (Actions) → Browser • Desktop • Playwright • PyAutoGUI │
│ 💾 Memory → Working + Episodic + Semantic (Redis • Qdrant) │
│ 📡 Event Bus → Kafka/Redpanda • Full audit trail & replay │
│ 📚 Learning → Extract skills & policies from successful runs │
└─────────────────────────────────────────────────────────────────────────┘
Why ARIA stands out:
| Pillar | What it means for you |
|---|---|
| Vision-First | VLM-powered UI understanding with multi-locator fallback — no brittle selectors only |
| Event-Sourced | Every step persisted; full audit trail and replay for debugging and compliance |
| Human-in-the-Loop | Safety gates for sensitive actions (login, CAPTCHA, payment) — production-safe by design |
| Local & Bilingual | Run entirely on your hardware; native Farsi STT (Whisper) and embeddings |
| Production-Ready | FastAPI + WebSocket API, Streamlit dashboard, Docker Compose, 81 tests |
Tech stack: LangGraph • Ollama / OpenAI • Qwen-VL • Playwright • Redpanda (Kafka) • Redis • Qdrant • Mem0
The Job Apply automation (LinkedIn, Indeed) is the first production plugin — the platform is built for more.
🔗 Explore ARIA → • 📖 Docs, ADRs, and MODELS.md inside the repo
📄 Governance-Safe Financial Document AI
Bilingual (EN/FA) • Quality Gates • Human-in-the-Loop Review • Replayable Lifecycle • Audit Endpoints
InvoiceMind is not an OCR benchmark or a generic prompt demo — it's a production-oriented platform for invoice extraction, human review, and governance-safe automation. Built for teams where traceability and control matter more than blind automation. Most invoice AI fails in production because decisions are hard to trust, explain, and control; InvoiceMind tackles that gap head-on.
End-to-end flow — from ingestion to final export, with explicit gates and audit at every step:
┌─────────────────────────────────────────────────────────────────────────┐
│ InvoiceMind — Pipeline & Lifecycle │
├─────────────────────────────────────────────────────────────────────────┤
│ 📥 Ingestion → Validation → OCR/Layout → LLM Extraction → Postprocess │
│ 📊 Routing (quality gates) → Review / Quarantine → Export + Audit │
├─────────────────────────────────────────────────────────────────────────┤
│ Run lifecycle: RECEIVED → VALIDATED → EXTRACTED → GATED → │
│ AUTO_APPROVED | NEEDS_REVIEW → FINALIZED │
│ Control: cancel • replay • quarantine (reason-coded) • audit/verify │
└─────────────────────────────────────────────────────────────────────────┘
Why InvoiceMind stands out:
| Pillar | What it means for you |
|---|---|
| Evidence-first | Policy and gate-based routing instead of confidence-only automation |
| Decision traceability | Every auto-approve or escalate tied to gate results and reason codes |
| Replayable & auditable | Full run lifecycle, cancel/replay, and audit endpoints for compliance and post-incident analysis |
| Local-first | Privacy-first inference; versioned config bundles and model registry (models.yaml) |
| Safe defaults | Quarantine and human review over aggressive auto-posting; NIST AI RMF & OWASP LLM–aligned |
Tech stack: Python 3.11+ • FastAPI • Next.js 16 • React 19 • TypeScript • SQLAlchemy • Alembic • SQLite • AGPL-3.0
ADR-001 (local-first), ADR-002 (evidence-first), ADR-003 (policy-driven gates) — design documented in the repo.
🔗 Explore InvoiceMind → • 📖 Docs, run.bat one-click startup, API surface in README
🧭 Evidence-First Web Extraction With Safe Repair Planning
Human Prompt → Validated Task → Read-Only Extraction → Structured Records → Evidence Critic → Repair Proposal
TrustTrace is a technical preview of an AI web extraction system that refuses to treat scraped output as trustworthy by default. It turns a human request into a validated task, performs bounded read-only extraction, exports structured artifacts, evaluates the result, and produces safe non-executing repair proposals when the output looks suspicious.
The key idea: web extraction should be auditable. Empty results, blocked targets, invalid URLs, contaminated parent-container text, and broken artifacts should be visible states — not silent failures hidden behind a spreadsheet.
Self-auditing extraction loop:
┌─────────────────────────────────────────────────────────────────────────┐
│ TrustTrace — Evidence Loop │
├─────────────────────────────────────────────────────────────────────────┤
│ 🧑 Human prompt → Natural-language extraction intent │
│ ✅ Task contract → Validated fields, target, limits, output format │
│ 🌐 Read-only run → Public-page observation and bounded extraction │
│ 📦 Artifacts → JSON / Excel / file manifests with hashes │
│ 🔎 Evidence Critic → Output quality, URL, artifact, and PDF checks │
│ 🛠 Repair Planner → Safe, non-executing, human-reviewable proposals │
└─────────────────────────────────────────────────────────────────────────┘
Why TrustTrace stands out:
| Pillar | What it means for you |
|---|---|
| Evidence-first | Extraction is followed by deterministic output checks before trust is assigned |
| Failure-visible | NO_RESULTS, NEEDS_REPAIR, and BLOCKED are explicit states, not hidden failures |
| Safe by design | No credential use, CAPTCHA/WAF bypass, stealth, proxy evasion, or unsafe retry behavior |
| Repair-aware | Suspicious output creates repair proposals, but repair execution remains human-reviewable |
| Portfolio-ready | Public snapshot includes sanitized samples, docs, UI demo, release notes, and safety model |
Tech stack: Python 3.11+ • Pydantic • Playwright boundary adapters • deterministic XLSX export • static bilingual UI
The public release includes a sanitized arXiv sample and a TrustTrace UI demo; raw evidence packs, videos, downloaded PDFs, local DBs, and generated run artifacts are intentionally excluded.
🔗 Explore TrustTrace → • 🚀 Release v0.1.0 • 🛡️ Safety model documented in the repo
🛡️ DriveShield — Real-Time Collision Risk Intelligence
End-to-end collision prediction platform using Nexar's BADAS-Open model.
- State-of-the-Art Prediction: Real-time risk analysis with vision models
- 100% Offline: Runs locally without external API calls
- Production-Ready: FastAPI backend + React TypeScript frontend
Tech: Python • FastAPI • React • TypeScript • PyTorch • Computer Vision
🔄 Hybrid Retail Recommender System
Production-ready hybrid recommender combining collaborative filtering & content-based ML.
- Results: 140% precision improvement, 175% recall improvement
- Scale: Tested on 38K+ user dataset
- Bilingual: English/Persian UI with RTL support
Tech: Python • FastAPI • React • TypeScript • scikit-learn • Docker
🌊 FlowCast — Surge Pricing & ETA Optimization Engine
Enterprise-grade intelligent pricing and ETA prediction for ride-hailing platforms.
- ETA Accuracy: +20% improvement over baseline
- Revenue: +10-25% efficiency per trip
- Price Stability: 30-40% volatility reduction
Tech: Python • FastAPI • React • GeoPandas • Time-Series Forecasting
💊 Pharmaceutical Supply Chain Agentic AI
Four-agent system for supply chain optimization using LangGraph orchestration.
- Logistics Costs: 40% reduction
- Stockouts: 67% reduction
- Forecast Accuracy: 95%+ (MAPE < 5%)
Tech: Python • FastAPI • LangGraph • Next.js • MongoDB • GPT-4o-mini
📚 More Projects
| Project | Description | Tech |
|---|---|---|
| Blood Cell Cancer Detection | CNN-based classifier with 99%+ accuracy | TensorFlow • Keras • Medical Imaging |
| Books Recommendation System | Production recommender, 8% sales increase | Collaborative Filtering • scikit-learn |
| Stock Price Collection | Automated data pipeline for finance ML | Web Scraping • Database Design |
| CIFAR-10 Classification | CNN image classifier, 90%+ accuracy | TensorFlow • Keras • CNN |
AI & LLM
ML & Data Science
Production & DevOps
Frontend
| Achievement | Description |
|---|---|
| 🥈 | 2nd Place — Tehran Provincial AI Competition (2022) |
| 🎓 | Member — Iran's National Elites Foundation |
| 📜 | Kaggle Notebooks Master |
| 📄 | Published Researcher — Health Science Reports (Wiley), ICVPR, AMLAI |
- M. Navaei et al. "Leveraging Machine Learning for Pediatric Appendicitis Diagnosis" — Health Science Reports (Wiley)
- M. Navaei, Z. Doogchi. "Machine Learning Models for Predicting Heart Failure" — ICVPR
- M. Navaei, M. Pahlevanzadeh. "Forecasting Forex Market Stock Prices Using Neural Networks" — AMLAI
| Role | Company | Period |
|---|---|---|
| Senior AI/ML Engineer | Daria Hamrah Paytakht | Jul 2024 – Present |
| Senior AI/ML Engineer | Educational Industries Research & Innovation Co | Nov 2023 – Jul 2024 |
| Data Science Team Lead | Diar-e Kohan CO. | Sep 2020 – May 2022 |
| Data Scientist | Diar-e Kohan CO. | Sep 2018 – Sep 2020 |
- 🚀 Building Agentic AI systems and LLM applications at innovative companies
- 💼 Production-grade AI systems that solve real business problems
- 🌍 Collaborating with international teams on cutting-edge AI/ML projects
- 🤝 Remote positions, contract work, or full-time opportunities worldwide
