Farbod Haeri

A team machine-shop build, and where I learned to actually make parts: a working scale cannon turned and milled from aluminum. We modeled every piece in SolidWorks, drew dimensioned and toleranced engineering drawings, then cut them on a manual lathe, a Tormach lathe, and Haas CNC mills before assembling and polishing the finished cannon.

Led mechanical design on a small team: built the chassis and drivetrain from scratch and iterated through three revisions to take first place in the competition. Everything was designed in CAD, 3D-printed and laser-cut, and rebuilt as fast as testing broke it.

Turning raw optical metrology into fatigue-life data for additively-manufactured steels. I built MATLAB pipelines that quantify defect fractions and surface roughness, designed and machined test fixtures, and ran fatigue tests across material variants while keeping the datasets clean and traceable.

Numerical Methods Toolkit · MATLAB

A from-scratch set of the core numerical methods in MATLAB: ODE solvers (Euler and my own RK45), a shooting method, numerical integration (trapezoid and midpoint), finite-difference derivatives, and root finding (bisection and fixed-point iteration).

StockAI · AI Trading Research System

A personal deep-dive into modern software and AI tooling: an automated trading research system that ingests public "smart money" signals (congressional trades, insider filings, 13F whales, activist positions), scores them, and paper-trades a risk-gated portfolio end to end, with backtesting behind every rule. The repo is private since it trades a live account, but I'm happy to walk through the architecture.

My senior capstone for a UCSD neuroscience lab: an automated touchscreen rig that trains group-housed mice on the 5-choice task. I handled the software side of it, which meant an operator dashboard (HTML/CSS/JS) with live per-mouse metrics, a Python/Flask "brain" that runs through every training stage in order, and the data-logging plus the Wi-Fi/MQTT link to the hardware.

A six-rotor drone whose rotors sit at fixed angles, so it can move in all six directions without a gimbal. It's meant for tracking a target and searching tight spaces like tunnels. I took it from CAD to a flight-ready build: the SolidWorks design, a 3D-printed monocoque airframe, thrust-stand testing of ducted vs. un-ducted props, and a 6-DoF flight controller that I tune in MATLAB and test in our motion-capture room.

A 1/10-scale car that surveys vegetation on its own. Press one button in a browser and it brings up centimetre-accurate GPS, drives a saved route with no driver, identifies trees, shrubs, and people from its camera as it moves, stops itself when the loop closes, and writes the report.

The detection runs on the car, on a 26 TOPS accelerator, not on a laptop and not in the cloud. I trained the model by having a large vision model label our own footage, then distilling that into a small one built for the chip. On hardware it holds 0.735 mAP50 at 41.8 frames per second, roughly forty times what the same model managed on the onboard CPU.