Erfan Pirjahandideh

I build ML experiments and the tools to review them.

I work on applied machine learning and full-stack tools. I'm especially interested in how brain-signal models perform on people they haven't seen before.

Open EEG viewer OpsDesk fixture Reachboard demo

Brain Signals Lab

Motor-imagery EEG on public PhysioNet data, scored on participants the model never trained on.

I kept each participant's recordings in a single fold. Across 50 people and five folds, CSP + LDA averaged 61.8% accuracy. A most-frequent-class baseline reached 50.4%. The viewer loads saved reports only—nothing retrains when you open a page.

Try it: open Participants and filter below chance to see spread across people.

Engineering results on one dataset—not a clinical or neuroscience claim.

Live viewer Code

Evaluation section of the Brain Signals Lab viewer with fold accuracy chart.
Saved evaluation report in the public viewer (screenshot source).

OpsDesk AI

A fictional support desk where proposals cite the handbook and a person approves before anything runs.

Changing an approved ticket proposal invalidates its approval. A second execution request returns the first receipt. I compared a rules baseline with a local language model on 60 fictional tickets: the rules produced 37 fully correct proposals and the model produced 25.

Try it: approve a plan change on the fixture, then repeat the run and watch duplicate handling.

The browser simulates the workflow; the Python service runs locally. All customer data is synthetic.

Live fixture Code

OpsDesk mobile fixture showing a receipt after simulated approval.
Mobile-width fixture replay (screenshot source).

Reachboard showcase

I built a fictional media-buying catalogue where the server owns cart totals and campaigns move draft → awaiting payment.

I seeded six packages with channel, placement, and RB credits prices. In the sample journey, add Echo Weekly (35) and Northline Sound (40), review cart → draft, and the locked campaign total is 75 credits. When I run Django locally, checkout is POST-only with CSRF and per-user reset—that backend is separate from the public browser simulation linked below.

Try the simulation: use the compare-row adds, review cart → draft, submit, then reset and shop again.

The screenshot is the verified public static fixture at draft (75 credits). A separate local Django catalogue capture shows the server-rendered app—not the Pages demo.

Browser simulation Code

Public Reachboard browser simulation at draft with Echo and Northline totalling 75 RB credits.
Public static fixture at draft (75 RB credits) after Echo + Northline (screenshot source). Live simulation linked above.

I used AI coding assistants on these portfolio repos; each project README describes scope and local-only pieces.