Work
Selected work
Palantir Foundry & AIP
Enterprise GenAI System — PIF
- Problem
- The PIF AI solutions team needed an enterprise-grade generative AI system covering retrieval, workflow automation, and governance.
- Approach
- Architected and hands-on built RAG across vector, hybrid, and graph-based retrieval; an ASR ingestion and transcription workflow; agentic workflows; and RBAC, lineage, and telemetry for governance.
- Outcome
- Shipped with LLM evaluation and regression-testing suites covering retrieval relevance, groundedness, latency, and cost — built with a team of Saudi fresh-graduate data engineers I led and mentored.
Ingest & transcribe (ASR)
Retrieve (vector / hybrid / graph)
Agentic workflows
Govern (RBAC, lineage, telemetry)
TONOMUS / NEOM
NEOM Enterprise Architecture
- Problem
- Multiple NEOM sectors — Food, Environment, Energy & Water, Hotels, and University of NEOM — needed enterprise AI and cloud architecture guidance, from first discovery to procurement.
- Approach
- Ran AI discovery workshops, then delivered enterprise AI and cloud reference architectures, landing zones, HLD/LLD, and NFR and cost/option analyses for each sector.
- Outcome
- Reference architectures and landing zones in place across NEOM sectors, with technology choices backed by procurement and technology advisory.
Discover (AI workshops)
Architect (reference arch., landing zones, HLD/LLD)
Validate (NFR, cost/option)
Advise & procure
UT Austin
Medical Imaging Diagnostic Pipeline
- Problem
- Brain-cancer classification models can be accurate yet opaque, making individual errors hard to diagnose.
- Approach
- Built a classification pipeline with Grad-CAM and SHAP explainability, plus a structured error-analysis process.
- Outcome
- Reduced false positives by 13%.
Scan input
Classify
Explain (Grad-CAM / SHAP)
Error analysis
UT Austin
CNN vs. Vision-Language Models for Explainable Medical Imaging
- Problem
- A CNN brain-tumor MRI classifier reached 98.4% accuracy in training and validation, but its remaining errors were hard to explain: heatmaps showed it attending to skull edges and anatomical boundaries instead of the tumor.
- Approach
- Analyzed the misclassified scans with Grad-CAM++, EigenCAM, and AblationCAM, then gave the same scans to LLaVA-Med, a biomedical vision-language model, using 2-shot and 4-shot in-context prompts, and compared its answers and explanations with the CNN's heatmaps.
- Outcome
- Of the 13 scans the CNN misclassified, LLaVA-Med classified 7 correctly with 2-shot prompts (53.8%) and 9 with 4-shot prompts (69.2%), explaining its reasoning in natural language.
Train & evaluate CNN
Explain errors (CAM methods)
Prepare few-shot examples
Classify with VLM
Compare explanations
UT Austin
A GPT Built and Trained from Scratch
- Problem
- A course assignment in Advances in Deep Generative Models: build a GPT-style language model end to end, with every component written from scratch.
- Approach
- Implemented a 30M-parameter decoder-only Transformer in PyTorch (6 layers, 6 attention heads, 256-token context) with causal self-attention, KV caching, and weight tying, plus a training pipeline with AdamW, a cosine learning-rate schedule, gradient accumulation, and mixed precision. Trained it on 50,000 TinyStories children's stories.
- Outcome
- Trained in about 5 minutes on a single RTX 5090 to a validation loss of 2.37, with no sign of overfitting. The model writes grammatical, locally coherent children's stories, and the ways it fails on out-of-domain prompts and extreme sampling settings show how autoregressive models behave.
Tokenize (GPT-2 BPE)
Transformer (6 layers, 30M parameters)
Train (AdamW, cosine LR)
Sample (temperature, top-k)
UT Austin
Predicting Apartment Rental Prices in Riyadh
- Problem
- Riyadh rental listings hide much of what drives price, such as private parking or a maid's room, inside free-text Arabic descriptions that structured fields do not capture.
- Approach
- Scraped 13,634 listings and cleaned them to 8,252 across 108 districts. A two-pass LLM pipeline extracted six amenity features from the Arabic descriptions with Qwen3, then verified them independently with ALLaM-7B, an Arabic-specialist model (over 94% agreement). Trained Ridge, Random Forest, and XGBoost models on the combined features and explained them with SHAP.
- Outcome
- Random Forest performed best on the test set: R² 0.735 and a mean absolute error of 9,492 SAR per year (15.9% MAPE). Private parking, a feature extracted by the LLM, ranked among the top five price drivers. Written up as a paper.
Scrape & clean listings
Extract features from Arabic text (LLM)
Verify with a second LLM
Train (Ridge, RF, XGBoost)
Explain (SHAP)
Side builds
AI Tender Evaluation
Document analysis and decision support for evaluating technical proposals submitted to an RFP.
View on GitHub Connected Vehicles In progress
Raspberry Pi + GSM/GPS + OBD-II telemetry to an IoT platform.