Abdulaziz
Khader
ML engineer who ships the whole stack — from enterprise AIOps agents at BMC down to a 360M-parameter LLM answering questions offline on an Arduino.
EDUCATION: UC San Diego — B.S. Computer Engineering '26, math minor
STATUS: OPEN TO NEW-GRAD SWE / ML ROLES
Quests cleared.
System Energy Efficiency Lab — Researcher
Optimizing LLMs to run on a new type of DRAM at UCSD's SEELab, working across model sizes from 1B to 7B parameters.
BMC Software — Software Engineering Intern
Built a new AIOps tool on the GenAI team folding customers' ServiceNow data into BMC's product line — LangChain/LangGraph CRUD tooling over ServiceNow's REST API, an ASGI server pushing real-time updates, and a semantic model that dynamically parses data configurations. Refactored 1000+ lines into a reusable tool factory that extends to products like Salesforce.
Product Manager Accelerator — Software Engineer Intern
Built a full-stack web app integrating reinforcement learning with a scalable database to evaluate players' chess games and surface move patterns and historical trends, wiring frontend workflows to Flask REST APIs, LangChain, and Google Gemini.
Ocelot — AI Studio Fellow
Built a sentiment-conditioned LLM for student mental-health conversations, validated through a 50-sample human evaluation. Optimized a 7B-parameter deployment to run in real time on local hardware, and compressed the training cycle through CUDA optimization across 5,600+ multi-domain samples.
Break Through Tech — AI & ML Fellow
Implemented machine learning and neural network models across ensemble regression, computer vision for digit classification, and NLP. Built a Deep Averaging Network for sentiment analysis on 2,000+ book reviews with a TF-IDF vectorizer, then benchmarked it against logistic regression and LSTM models using AUC ROC.
Think Neuro, Inc. — Summer Research Intern
Researched how far neurofeedback can enhance physical rehabilitation for patients with weak or no motor function. Extracted 100+ papers from Web of Science and analyzed them in R, producing bibliometric and sentiment graphs, then presented a poster of the findings at the Think Neuro Summer Symposium.
UC San Diego — Instructional Assistant
Mentored 800+ students through computer systems and software tools — C pointers, dynamic memory allocation, and bitwise operations — tutoring one-on-one, assembling assignments, and proctoring and grading exams.
Triton Robocup — Mechatronics Engineer
Embedded systems development in C on a 32-bit ARM Cortex-M STM32. Programmed and debugged electronic speed controllers for brushless DC motor control, UART with interrupts for Raspberry Pi data, I2C for IMU reads, and CAN bus for motor commands and encoder feedback. Sensor-fused accelerometer and gyroscope data with a Kalman filter to kill drift.
The big ones.
RingDown
A voice-driven ambulance routing system that cuts hospital offload delays: a paramedic speaks a patient report into a handset, faster-whisper transcribes and spaCy redacts names, addresses and birth dates on the laptop, and only redacted text reaches the decision model, whose confidence bands route each report to act, one clarifying question, or a human hand-off. Hospital choice stays in plain code — capability filter, Mapbox drive time, and expected wait counting ambulances already headed there — before Twilio phones the charge nurse with a spoken heads-up. Across 48 scripted reports it handled 79% alone at 100% accuracy and a 605 ms median. I built the voice path and the medic handset.
Remem
A wearable visual-memory search engine: a forehead-worn iPhone streams JPEG frames to a laptop, where EmbeddingGemma 2 embeds every frame and every question into one 768-dimensional space, so asking where your keys are returns the newest frame they appear in and how long ago it was seen. Capture through ingest was mine — live frames, ffmpeg MP4 replay, and EXIF-dated camera-roll imports landing in one JPEG-plus-SQLite store behind a bounded queue that drops the oldest frame, never the newest. Every frame is indexed and searched on the laptop: no cloud APIs, no audio recorded.
Building Blocks
A real-time collaborative workspace for system design: React Flow over Liveblocks gives a shared canvas with live cursors, presence and multiplayer editing, and an AI agent turns a plain-English prompt into nodes and edges written straight into the room. Generation runs as a durable Trigger.dev background task so a long agent call survives a reload, Clerk handles auth, and Prisma over Postgres holds projects, collaborators and the Markdown specs the finished graph compiles into. Ships with a starter library of monolith, microservice, event-driven and serverless templates.
RoboRanger
An edge AI wildlife identification robot on an Arduino Uno Q: a MobileNetV3-Large TFLite classifier (75%+ accuracy across 51 species, calibrated with temperature scaling and TTA) paired with a quantized SmolLM2-360M via llama.cpp for fully offline inference. RAG over sqlite-vec with cosine-similarity thresholding gates hallucinations, and a networked mode streams to an ElevenLabs agent for conversational narration. Won against 100+ teams.
GPS IMU Fusion
A 5-state Extended Kalman Filter fusing Adafruit BNO08X IMU data with u-blox C099-F9P GPS to eliminate dead-reckoning drift on a self-driving car. Refactored the vehicle's sensor pipeline into vectorized tuples and built backward-compatibility layers so legacy MPU6050/9250 sensors route into the new EKF without crashing the vehicle loop. Shipped upstream into DonkeyCar, the open-source self-driving platform.
YouPresent
A Unity VR project for building public-speaking confidence by simulating a conference stage. Google Gemini reads the presenter's gaze and hand motions to generate coaching feedback, an Adobe Mixamo audience applauds at the end, and a recording camera lets you rewatch yourself as your own avatar from the audience's seat.
Toolbox.
Not everything is shippable.
Perpetually mid-novel. The queue only grows.
Plays them. Builds them. Hence this whole page.
Weekend fixture. Midfield, mostly honest tackles.
String tricks. You met the physics version up top.
Yes, really. Ask for a demo in the interview.
Let’s talk.
One email gets you an engineer who documents everything, demos on time — and yes, still brings the yo-yo.
akhader@ucsd.edu© 2026 ABDULAZIZ KHADER · HANDCRAFTED IN SAN DIEGO