I build production AI systems that are observable, testable, evidence-driven, and safe to automate.

Senior AI/ML Engineer with 7+ years across applied machine learning, LLM/RAG systems, agentic architectures, AI/ML platforms, and production automation. My strongest work sits at the boundary between AI research, software engineering, and operational reliability.

Currently a Senior Machine Learning Engineer at Picnic on a part-time, remote basis.

My sweet spot: turning a promising AI capability into a system that can be reviewed, tested, shipped, operated, debugged, and trusted.

A production-oriented desktop/web product for non-EU candidates who need visa-sponsorship and relocation evidence, not another generic job board.

• Scheduled ingestion from public employer ATS sources with source-health and retry semantics.
• Deterministic sponsorship / eligibility logic with hard-negative handling and evidence provenance.
• Versioned, hash-verified market catalogs with transactional updates and local candidate state.
• FastAPI + Next.js, PostgreSQL/SQLite, browser tests, dependency audits, coverage gates, Windows packaging, and release checks.
• English/Persian UX with real RTL support.

Signal: production ownership · data pipelines · product ML · deterministic decisioning · CI/CD · release engineering

An open-source reliability toolkit for preserving agent failures as portable, versioned test artifacts.

Capture → AgentCase → Mock Replay → Diff → Regression Test

• Framework-neutral execution capture with redaction and atomic local persistence.
• Fail-closed replay that never silently falls back to live model or tool execution.
• Exact, structural, and normalized deterministic diffs with pytest regression integration.
• Explicit provider instrumentation boundaries, including OpenAI Python SDK integration.
• Contract, unit, integration, security, replay, diff, and regression coverage.

Signal: agent reliability · reproducibility · developer tooling · safety boundaries · regression engineering

A contract-first platform for observable, replay-aware, human-supervised agentic automation.

• Separates Brain, Eye, Hand, Memory, Safety, Replay, Learning, Voice, and LLM concerns into explicit system boundaries.
• Uses structured orchestration instead of hiding control flow inside prompt chains.
• Supports local models, vision-assisted UI understanding, tool execution, HITL routing, memory, and durable traces.
• Includes unit, integration, and E2E test structure plus build/test/docs workflows.
• Public repository is a curated implementation snapshot rather than a claim that every internal capability is public.

Stack: Python · LangGraph · FastAPI · Redis · Qdrant · Playwright · VLM/OCR · Docker

A local-first invoice-processing platform designed around quality gates, human review, auditability, and replayable decisions rather than blind OCR automation.

• End-to-end ingestion, validation, OCR/layout, LLM extraction, post-processing, routing, review, and export.
• Policy-driven gates with reason-coded escalation and quarantine states.
• Replay/cancel lifecycle, audit endpoints, versioned runtime configuration, and model registry.
• Bilingual FastAPI + Next.js/TypeScript product surface.
• CI covers backend unit/integration tests, frontend type checks/tests, dependency audits, and performance smoke scenarios.

Signal: local LLMs · document AI · governance · HITL · traceability · production workflows

A technical preview of evidence-driven web extraction that checks its own output instead of treating every scrape as success.

Prompt → Validated Task → Bounded Extraction → Evidence Critic → Safe Repair Proposal

• Detects suspicious outputs such as zero records, field mismatch, invalid URLs, parent-container contamination, and blocked targets.
• Distinguishes success, repair-needed, and blocked states instead of collapsing them into empty results.
• Produces non-executing, human-reviewable repair proposals.
• Includes sanitized success/corruption/blocked demos and focused evaluator, executor, and repair tests.
• Intentionally refuses unsafe access-control bypass and blind retry behavior.

Signal: evidence-driven automation · agent evaluation · failure semantics · safe repair planning · human oversight

A reproducible rare-event ML study using first-day EHR data across 119,743 encounters.

• Leakage-controlled feature construction with a pre-specified frozen held-out evaluation boundary.
• XGBoost primary model under extreme class imbalance.
• Average Precision, AUROC, Brier score, calibration, uncertainty, subgroup analysis, and fixed alert-budget evaluation.
• Public release includes code, frozen configuration, tests, provenance notes, and aggregate reproducibility artifacts without exposing source patient data.
• Weaknesses and limitations remain visible instead of being tuned away after test evaluation.

Signal: applied ML research · imbalanced learning · evaluation design · clinical AI · reproducibility

• Tool-Memory-Evidence Retrieval Benchmark — dense/sparse/reranked/hybrid retrieval evaluation for agent tools, memory, and evidence with Qdrant + FastEmbed.
• Google Scholar Scraper V2 — local Windows research utility with deterministic tests, explicit failure states, installer/portable builds, checksums, and release hardening.
• DriveShield — local/offline collision-risk intelligence with BADAS-Open integration, FastAPI, React/TypeScript, and evaluation pipelines.
• FlowCast — surge-pricing and ETA optimization with demand forecasting, geospatial routing, APIs, and a realtime dashboard.
• Pharmaceutical Supply Chain Agentic AI — multi-agent forecasting, routing, inventory, and monitoring workflows.
• Hybrid Retail Recommender — collaborative + content-based recommendation with sparse matrices, cold-start handling, offline ranking metrics, API, UI, and Docker.

Older learning-oriented repositories remain public, but are intentionally not part of the primary showcase.

AI / ML
Python · PyTorch · scikit-learn · XGBoost · Transformers · OpenAI APIs · Ollama · local LLMs/VLMs · RAG · embeddings · reranking · CV · NLP · time series

Agent & Retrieval Systems
LangGraph · tool orchestration · memory systems · Qdrant · FastEmbed · human-in-the-loop workflows · agent evaluation · replay/regression tooling

Platforms & Data
FastAPI · PostgreSQL · SQLite · Redis · Kafka/Redpanda · SQLAlchemy · Alembic · MLflow · data-quality and drift monitoring

Production & Delivery
Docker · GitHub Actions · CI/CD · pytest · mypy · Ruff · dependency/security checks · release automation · Windows packaging

Product Engineering
Next.js · React · TypeScript · REST/WebSocket APIs · bilingual English/Persian interfaces · RTL

My work also includes applied machine learning research in healthcare and forecasting.

• 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.

ORCID →

I am most interested in roles where models are only one part of the system and engineering quality matters just as much as model quality:

Senior AI/ML Engineer · Generative AI / LLM Engineer · Agentic AI Engineer · AI/ML Platform Engineer · MLOps / LLMOps · Applied AI Engineer · Applied Scientist · Senior Data Scientist

Open to international teams, remote collaboration, and relocation opportunities in Europe with visa sponsorship.

Let's build AI systems that survive contact with production.