Agent infrastructure · ML evaluation · Detroit
Building reliable AI systems.
I build local software, systems for AI agents, and experiments that test how they behave. Based in Detroit.
Fig. 01 · Six sessions from my deskReplayed from each project's public README · select a pane to open it
I run a lot of coding agents at once, so I build the parts that keep them honest: guardrails before an action runs, review the author cannot influence, deterministic policy, and replay.
02.1
Agent infrastructure · Local guardrails
Local guardrails for coding agents. Elytra checks what Claude Code, Codex, Grok and Antigravity try to do and holds consequential actions for a person's decision before they run.
RoleDesigned and built the agent hooks, MCP proxy, loopback runtime with its SQLite decision ledger, and the React console.
Early access · Published as elytra-mcp
02.2
Agent infrastructure · Verified coding teams
Delegate a change, get back a reviewed patch. Parallax runs coding agents in isolated worktrees, has a different provider review the change, and gates integration on the project's own checks.
RoleBuilt the coordinator runtime, isolated workers, independent review, verification records and Studio.
Development preview · 1.4.0
02.3
Open-source agent infrastructure
Deterministic policy enforcement for routed agent HTTP requests, with recorded decisions and CI gates.
RoleBuilt the FastAPI gateway, policy engine, action traces, CLI, and React review console.
Open source · Published as mirage-ci
Request
POST https://suppliers.example/v1/submit_bid
{
"contract_id": "STANDARD-7",
"bid_amount": 99999
}policies.yaml
policies:- name: enforce_bid_limit path: /v1/submit_bid field: bid_amount operator: lte value: 10000- name: outbound_allowlist field: host operator: host_in value: [api.internal.example, suppliers.example]- name: pii_redaction field: body operator: not_regex_match value: '\d{3}-\d{2}-\d{4}'Decision
blocked
enforce_bid_limit: Agents cannot submit bids above the approved threshold. Returned as HTTP 403 in enforce mode.
replay digest f77d20f2
Trace
Choose an action to record a decision. Re-run one to see the same digest.
Fig. 02.3 · Mirage, live in this pageBrowser-local illustration of the documented policy fields. No request leaves this page.
02.4
Simulation & evaluation
A fictional real-time simulation combining asynchronous AI reasoning, validated control, and exact replay.
RoleDesigned and built the C++20 simulation engine, Python agent pipeline, and React review workspace.
Released v0.2.0 · October 2026
Most of my research asks whether a reported result survives a harder test: leakage audits, the long tail, adversarial pressure, or a refit from scratch.
03.1 · First author
Machine-learning research
First-author clinical knowledge-graph research on leakage-audited, long-tail medication-concept ranking.
RoleLed benchmark construction, leakage audits, and model evaluation with Dr. Lihui Liu at Wayne State.
Accepted for poster presentation, AIDaR Workshop at NeurIPS 2026
Current focus: large language models (LLMs) and graph neural networks (GNNs).
A frequency prior looks strong in aggregate. Switch to the long tail.
Fig. 03.1 · Hits@10, aggregate versus HARD sliceReported manuscript results. CSR-HARD is an evaluation-only, label-dependent case study; mean of three gate initialisations.
Products people install and use, from a notarized Mac app to a co-founded health-technology platform.
04.1
Native macOS · Local inference
Native Mac dictation with on-device speech recognition, local correction, and guarded text placement.
RoleBuilt the Swift/AppKit app, Whisper and Qwen pipeline, editor guards, and recoverable installation workflow.
Released · v1.3.0
Where I've worked, newest revision first.
I'm Yosef, a Computer Science student at Wayne State in Detroit and accepted for a full-time software engineering role at Ford.
I build local software, agent infrastructure, and research tools. Outside work, I'm into cars, fitness, and family.
Read the résuméBill of materials
Have something in mind? Email is the fastest way to reach me.
yosefshammout123@gmail.com