I build AI systems for markets and classrooms.

An AI back office that a human still signs off on

In summer 2026 I was an AI Automation & Agents Engineering Intern at Hwy Haul, where I built FreightDesk AI / LoadLink — an AI freight back office running entirely on-premise.

A small brokerage's AR team spends most of the day retyping the same messages: POD requests, invoice disputes, W9 chases, past-due nudges, all while reconciling several disconnected systems by hand. The system takes WhatsApp and email intake, extracts structured load details out of free text, and drafts a reply in the brokerage's own voice — then holds it for one-click human approval. FastAPI, a local LLM, a green automated regression suite, the full set of finance and accounts-receivable pain points, zero third-party API cost, and no data leaving the building.

The lesson wasn't the model. Any competent local LLM handles this. The reliability lived in the boring parts: process management, retry logic, VPN-aware degradation, and an always-reply design so no carrier is left waiting on an ambiguous “SJ/LA?”. The right question about an AI product isn't which model — it's whether the people in the loop trust it enough to hit Send.

Does the language you think in change how you learn to code?

I'm a co-author on an ACM-published SIGCSE Technical Symposium 2026 poster abstract, Exploring Bilingual Coding for Inclusive Computer Science Learning, and on a paper accepted to the Papers track at SIGCSE Virtual 2026, Does Language Matter in Programming? Both are out of Dr. Ethel Tshukudu's lab at SJSU.

The underlying study ran coding sessions with 60 bilingual San José State participants on the multilingual Hedy platform, measuring confidence, computing identity, enjoyment, and motivation before and after, with novice participants showing the largest confidence gains. See the published poster abstract and the accepted SIGCSE Virtual 2026 paper for the full methodology and results — I'm not restating a topline percentage here that isn't in the published record.

The number isn't the interesting part. What stayed with me showed up in the qualitative data: students who had spent years silently translating every keyword before typing it described the bilingual sessions as the first time coding felt like thinking instead of translating. That friction is real, and it's measurable.

What research actually taught me had little to do with language. It was how much of a claim I was willing to defend in front of a reviewer who didn't care how I felt about it. Every effect size needed an interval beside it. That bar moved and stayed moved — which is why the disclosures below retract one of my own earlier trading results.

Everything else

An NLP sentiment-scoring prototype for equities, built on IBM Watson through IBM SkillsBuild. A local AI financial analyst running on Qwen, wired into iMessage, with no cloud dependency. The Silicon Valley Entrepreneurship Club website. I trade equities, options, and futures on the side, and I publish backtest confidence intervals on my own strategies — including the ones that didn't survive re-testing.

I work in Arabic and English, and I'm based in San Jose, California. Every project here was built solo. The stack, the constraints, the tests, and the limits are all documented, because a capability list is easier to write and worth much less.

Start with the Hwy Haul case study for the clearest implementation detail, then the research abstract, then the projects.