sec/00 — identity
AI Systems Engineer — Agentic Infrastructure & Applied ML
I build the infrastructure underneath AI systems — from a vector database written from scratch in C, to multi-agent frameworks that debug production ERP migrations. Based in Rasht, Iran.
STATUS: open to select engagements
sec/01 — stack
Technical range spanning systems programming, applied ML, and the infrastructure that connects them.
sec/02 — projects
Production systems, research tooling, and open-source infrastructure.
Self-hosted ERP migration for a live production system
A self-hosted pipeline migrating a full Odoo 17 Enterprise deployment to version 19 — dual-hop schema migration, ORM-lifecycle-aware data transforms, and a 6-layer validation framework spanning structural integrity through a full HTTP surface crawl. Debugged by a 13-agent AI framework with persistent memory across sessions.
Persistent structured memory for AI debugging agents
A SQLite-backed knowledge graph and CLI giving opencode debugging agents durable memory across sessions — typed nodes (issue, hypothesis, root_cause, repair_plan...) and a pure-heuristic dispatcher that routes work through an eight-stage triage-to-postmortem workflow.
Runtime behavioral auditor for MCP servers
An strace-based tool that catches the gap between what an MCP server's tool descriptions claim and what its syscalls actually do. Supports stdio, SSE, and streamable-HTTP transports, with JSON output and a --fail-on flag for CI.
A vector database written from scratch in C11
HNSW and Vamana indexes built from scratch in C11, benchmarked against FAISS on SIFT1M and GIST1M with honestly-reported crossover points. Includes a hand-written ARM NEON ternary GEMM kernel (BitNet b1.58-style) reaching 40+ GOP/s.
Generating runnable agentic systems from a prompt
A local-first framework built on LangGraph that turns a natural-language prompt into a runnable multi-agent system. Designed around one architectural thesis: a deterministic backstop over prompt hope.
sec/03 — research note
Exploring the intersection of consciousness, AI, and reality.
Core thesis: even if AI systems achieve consciousness, their — and our — grasp of reality remains incomplete, bounded by Gödel's incompleteness theorem. Adaptive minds, however, engage in meta-evolutionary processes, constructing ever-richer representations through the recursive pursuit of understanding.
This work explores how consciousness transcends representational limits through recursive self-modification, introducing a "Mesh of Realities" framework where reality consists of interconnected layers of formal systems, each with inherent incompleteness — and a "Trans-Reality Axiom Exchange" model where consciousness serves as a bridge-builder between those layers.
A formal model of how minds transcend representational limits through recursive self-modification.
Reality as interconnected layers of formal systems, each with inherent incompleteness.
Consciousness as a bridge-builder between formal layers of reality.
Dynamic axiom discovery across the mesh enables transcendence of local incompleteness.
"Consciousness is not the static possession of knowledge, but the infinite, recursive pursuit of understanding — a process that reveals both the inexhaustibility of reality and the creative power of intelligent minds."
sec/04 — writing
Notes on agentic engineering, systems programming, and the philosophy of mind.
Why are there patterns everywhere in the universe — and what happens when the question folds back on itself?
Every article I'd written until now was AI-generated. This one isn't — a look at how I nearly lost ownership of a 10,000-line Odoo migration codebase, and the one habit that got it back.
Debugging a two-hop Odoo 17→19 migration with a thirteen-agent AI framework kept hitting the same wall: every session started from zero. A SQLite knowledge graph and CLI — plus a hard-learned lesson about what "durable" actually requires — fixed that.
A research note on formalizing the intuition of a self-producing world. If reality is a structure that generates itself, that's a fixed point of a generating operator — and mathematics already has theorems about when those exist and are unique.
Tool descriptions are marketing copy — they don't tell you what a server actually does at the syscall level. Why I built mcp-behave, an strace-based runtime auditor, and what it turned up.
HNSW, Vamana, and a hand-written ARM NEON kernel: what it took to get vecdbc competitive with FAISS, where it actually wins, and where it still loses.
sec/05 — contact
Open to opportunities in agentic AI infrastructure, systems-level ML, and research collaboration.
navid72m@gmail.com
IRST (UTC+3:30) — Rasht, Iran
I'm always interested in discussing new opportunities, research collaborations, or conversations about agentic AI infrastructure. Whether you're looking for a systems-level ML engineer or someone who can bridge theory and practice, let's connect.