DHAHAB · Projects

Everything I've actually built.

Yusuf Gadelrab is a computer science student at San José State University (BS Computer Science, expected May 2028) who ships in public. Each project below has its own case study page — what it does, how it was built, and, where it applies, the result that went against him. Nothing here is a mockup with a made-up metric attached.

The academic side lives on the research page. This page is the engineering.

14 Builds on this page, each with its own case study Counted below · no entry without a page
208 Passing tests in DIRA, the zero-dependency scanner pytest · MIT · open on GitHub
4,933 Trades in the walk-forward audit that invalidated most of the setups 129 symbols · 10 years · no lookahead
$0 Live capital ever traded — the promotion gate has never opened Execution engine, paper only

Two of those four are zeros, and they are on the page for the same reason the other two are: a number you cannot check is decoration. Each resolves to the case study it came from.

Live shipped and in use
Prototype built, not in production
Local-first runs on local models, no API bill

Reply within two business days · you deal with me directly, not an agency


01 · AI & Automation

Systems that do the boring work

Built on local models wherever possible — no per-token bill, and the customer's data never leaves their machine.

Founding pilots open

FreightDesk AI

An AI back-office clerk for small freight brokerages. It reads the billing and accounts-receivable inbox, works out what each message actually is — a rate dispute, a missing BOL, a chase on an unpaid invoice — and drafts the reply, all on a local model so no customer data leaves the brokerage. Small brokerages lose hours a day to that inbox and cannot justify a full-time clerk for it; this is the shape of the problem an LLM is genuinely good at, provided you keep a human on the send button. Sold as a founding-client pilot at $999 setup plus $249/month, with a 3-day install.

PythonFastAPILocal LLMOllama
Read the case study →
Internship build

HwyHaul LoadLink

An AI intake and accounts-receivable automation built during a freight-brokerage internship, covering its finance workflows end to end. The design constraint that shaped everything: it had to run entirely on local models, at $0 in API cost, because freight paperwork is full of customer and carrier data that should not be posted to a third-party endpoint. a passing regression suite, because a finance workflow that is wrong 5% of the time is worse than no automation at all.

PythonLocal LLMRegression suite$0 API cost
Read the case study →
Open source · MIT

DIRA — security scanner

A security audit for startup codebases in one command: hardcoded secrets, dependency CVEs pulled from OSV.dev, configuration and infrastructure-as-code misconfigurations, dependency license risk, secrets buried in git history, and live TLS and header checks — ending in a startup security-readiness score. Written in Python with zero dependencies on purpose, because a security tool that drags in fifty transitive packages is expanding the attack surface it was hired to shrink. 208 passing tests, MIT licensed, with SARIF and SBOM output for CI.

PythonZero-dependencySARIFCycloneDX208 tests
Read the case study →
Prototype

EventReels

A local-first ffmpeg pipeline that takes raw event footage and returns a finished 9:16 highlight reel — cut detection, beat-aligned edits, captions, export — plus a browser studio to drive the whole thing without touching a command line. The interesting engineering is in treating ffmpeg as a compiler target: the studio produces an edit plan, the pipeline renders it deterministically, and the same input always yields the same reel.

PythonffmpegWeb studio
Read the case study →

02 · Trading research

Software built to disprove itself

Published with the results that failed. No live capital, no profit claims, none of it financial advice.

Live · results published

Swing Screener

A swing-setup screener sitting on top of a backtest engine written specifically to make lookahead bias impossible — signals can only ever see data that existed at signal time. Then it was pointed at its own headline result. Adversarial walk-forward validation invalidated most of the setups being traded: only the anchored-VWAP reclaim survived. Its original headline figure, measured on a far smaller sample, then failed a second and harder re-test and was replaced by +0.117R over 4,933 trades on a 129-symbol, 10-year universe (95% CI +0.057 to +0.174) — and a risk-matched random entry already captures +0.086R of that, leaving the signal's own contribution near +0.030R with a confidence interval crossing zero. VCP came out breakeven; a gap and opening-range-breakout proxy came out at −0.28R and was dropped. The case study is the audit, not the strategy.

PythonWalk-forwardNo-lookahead
Read the case study →
Paper only

Broker-agnostic execution engine

An execution layer that treats a broker as a swappable adapter rather than the architecture. Deterministic order IDs and a claimed-before-every-call journal mean a crash mid-order is recoverable instead of ambiguous; reconciliation is idempotent, so replaying it can never double-fill. The risk layer fails closed by default, and a hard promotion gate refuses to connect a live broker until a paper track record clears a positive-expectancy threshold. 160 passing tests. It has never traded real money, and the gate is the reason.

PythonIdempotent ordersFail-closed160 tests
Read the case study →

Stated plainly, because it is the part that matters: the anchored-VWAP reclaim is the only setup that survived walk-forward, and it is not a validated edge. A risk-matched random entry already captures +0.086R of the +0.117R, which leaves the signal's own contribution around +0.030R with a confidence interval that crosses zero. No live capital has ever been deployed behind any of it, and none of it is financial advice.


03 · Product & design

Prototypes with an argument

Each one exists to test a specific product thesis, not to pad a portfolio.

Prototype

EcoImpact

A community trash map paired with a quantified impact meter, so a neighbourhood cleanup produces a number instead of a vibe. FastAPI, SQLite and Leaflet, designed privacy-first from the schema up — location reports are the whole product, which makes deciding what not to store the central design decision rather than an afterthought.

FastAPISQLiteLeaflet
Read the case study →
Concept

SpartanEats

A student-run food delivery concept for San José State: campus-priced delivery from local spots, built by SJSU students for Spartans. The thesis is unit economics — the national delivery apps take a margin that only works at their scale, and a single dense campus with student couriers is exactly the market where a smaller operator can undercut them.

Product conceptMarketplace
Read the concept →
Prototype

AU

A social feed with the outrage machinery deliberately removed: strictly chronological, no vanity metrics, no quote-posts, no verification hierarchy. Every one of those features exists to increase engagement, and every one of them is also the mechanism by which timelines get worse — so the prototype is an argument that you can subtract them and still have a usable product.

Product designFront-end
See the prototype →
Design study

AURUM

A luxury super-app study: five product surfaces — retail, stays, cinema, dining and motors — running on one design system and one faceted search engine. The point is the shared substrate. Five verticals that look like five apps but are one component library and one query layer is the hardest thing to get right in a multi-surface product, and the easiest thing to fake in a mockup.

Design systemFaceted searchJavaScript
Open the study →

04 · Free tools

Client-side, no signup, no backend

Every one is a static page. No accounts, no telemetry — data stays in your browser's local storage.

Live · free

Visa Navigator

A change-of-status and CPT/OPT planner for international students: a backward-planned I-539 timeline, conflict flags for the dates that quietly invalidate each other, and per-stage checklists. Built because immigration timing is a scheduling problem being solved by anxious people at 2am with twelve forum tabs open, and it runs entirely in the browser precisely because nobody should upload their immigration situation to a stranger's server.

Client-sideNo backend
Open the planner →
Live · free

Apply OS

A private tracker for internship, scholarship, grant and fellowship applications: a deadline radar that surfaces what is about to close, a kanban board for stages, and recommender tracking so a letter request never silently goes stale. It is the same system used on a real application season, not a demo of one.

Client-sideKanban
Open the tracker →
Live · free

Freight broker calculators

Six calculators freight brokers otherwise rebuild in a spreadsheet every quarter: rate per mile, deadhead cost, broker margin, detention, fuel surcharge, and break-even. No signup, runs in the browser, and each one shows the formula so it can be checked rather than trusted.

Client-sideFreight
Open the calculators →
Live · free

Risk toolkit

Position sizer, R-multiple and expectancy calculators, a Monte Carlo risk-of-ruin simulator, Kelly fraction, and break-even win rate. Built out of the trading research above, where the recurring lesson was that position sizing decides survival long before setup selection decides returns.

Client-sideMonte Carlo
Open the toolkit →
Live · installable

Offline PWAs — TradeLog, Cut, Fire, GRAMPA

Four installable progressive web apps with no backend at all: TradeLog (an R-multiple trading journal), Cut (a weight tracker), Fire (a daily ritual checklist), and GRAMPA (a big-button phone helper for older adults). They install to a home screen, work with no connection, and store everything in local storage on the device. The full apps and projects index lives here too.

PWAOfflinelocalStorage
Browse the apps index →

05 · Elsewhere

The rest of the work


FAQ

Before you click away

The three questions this page exists to answer.

Which of these is actually finished?

Check the status chip on each card. Live means shipped and in use today — DIRA, Swing Screener, the SIGCSE TS 2026 poster abstract, and the four free client-side tools. FreightDesk AI is shipped and priced with founding-client pilots open, but no client is installed yet. Local-first means built and tested end to end but running on local models only — HwyHaul LoadLink and the trading bot. Prototype or Concept means it works but has not shipped to a real user yet — EventReels, EcoImpact, SpartanEats, AU, and AURUM.

Did he actually build these, or assemble them from templates?

Check the repo and the test count, not the claim. DIRA is MIT-licensed, zero-dependency Python with 208 passing tests, open on GitHub. HwyHaul LoadLink shipped the AR and billing workflow load with a passing regression suite during a real freight-brokerage internship. The trading bot has 160 passing tests and a promotion gate that has never let it touch real money. Every project links to its own repo or case study.

What am I supposed to do with this page?

Click into whichever project matches what you need to see — each one links to a full case study. If it's about hiring him or building something like one of these, use Work with me below. If it's just to know when something new ships, leave an email under New Builds & Releases — nothing else, no sales sequence.


Stay in the loop

New Builds & Releases

One email when something new on this page ships. Nothing else, ever.

One email when something new ships. Nothing else, ever.


Hiring, or want one of these built for you?

You deal with him directly, not an agency, and the reply lands within two business days — including the honest no when something isn't a fit.