Computational Linguist & Research Engineer

I bridge computational linguistics research and production systems that ship—and last.

The intersection is rare: NeurIPS/NAACL publications, tools still running at Apple eight years later, and a path from Khartoum's #1 exam scorer to debugging dialectal Arabic ASR at MIT to surviving Georgia Tech's masters program. I write about what I learn along the way.

Writing recognized by: Jeff Dean · Reddit AI Community (200K+ views)

Read My Writing →

Scroll

Systems Engineering

Constraint-driven architecture for real-world deployment

2025 · Systems Engineering

Local-First Voice Agent: High-Latency Orchestration on Legacy Hardware

OpenAI Whisper Ollama Mistral 7B Python Gradio

The Constraint: 2015 MacBook Pro with no GPU (integrated graphics only), limited RAM, thermal throttling. Cloud APIs are easy. Local orchestration on legacy hardware is the engineering challenge.

What Cursor Fixed: NumPy version conflicts breaking PyTorch, insecure subprocess calls, missing error handling, wrong model names. It caught almost everything.

What Cursor Missed: After all fixes, I tested: "What are the human rights principles of the UN?" Whisper transcribed it. Ollama hung. The problem: Mistral needed >30 seconds on my hardware. Cursor assumed my machine was fast. Solution: raised timeout to 5 minutes. AI assistants generate code. Engineers debug systems.

Why This Matters

Demonstrates systems thinking over tool usage. Most LLM demos assume cloud infrastructure or beefy GPUs. Real-world deployment means legacy hardware, air-gapped environments, and privacy-first constraints. This optimizes for the 90% case, not the ideal 10%.

2020 → 2025 · Architecture Analysis

Evolution of NLP Architectures: Benchmarking the 5-Year Collapse of Static Embeddings

Python spaCy scikit-learn DBSCAN LLMs

The Question: How did the NLP pipeline collapse between 2020 and 2025? This project analyzes the architectural shift from template-based synthetic data (spaCy + GloVe 2014) → LLM-enhanced generation (Transformers + zero-shot prompts) in disease outbreak detection.

2025 Measured Results: 168 locations (84%), 154 diseases (77%) extracted from 200 LLM-generated headlines. 13 spatial clusters identified via DBSCAN. The architectural shift: better generalization (handles post-2020 entities), but new tradeoffs—API costs, non-determinism, explainability loss.

What Didn't Change: DBSCAN clustering. Why? It's geometry, not NLP. The 2020 geospatial code is still optimal in 2025. Senior engineers preserve working components during paradigm shifts—they identify what doesn't need LLMs.

Why This Matters

Quantified paradigm shift analysis, not just "LLMs are better." Documents the tradeoffs textbooks skip: when would you NOT use 2025's approach? Latency-critical systems, budget constraints, regulated industries requiring explainability. Juniors chase trends. Seniors benchmark the cost.

The Story

From Khartoum to Georgia Tech to wherever this goes next.

I ranked first among 400,000 students on Sudan's national exams in 2011. That number meant something there. It meant government recognition, newspaper interviews, a kind of celebrity status that's hard to explain to Americans. It also meant expectations I'd spend years reconciling with reality.

Jackson State University gave me my undergraduate foundation (4.0 GPA, the first student there to achieve it). Georgia Tech brought me to the intersection of linguistics and computation. I published at NeurIPS and NAACL workshops under Jacob Eisenstein, applying kernel methods to geolinguistic analysis (mining dialectal patterns from Lyon Twitter data). MIT CSAIL had me building Arabic speech recognition systems under James Glass. And at Apple Siri, I built a bug logging tool that streamlined debugging 10x for engineers, was selected for Apple's internal research summit, and is still in use today, eight years later.

In September 2025, after 8.5 years, my asylum case was finally approved. Eight and a half years of uncertainty, of building a life on temporary foundations. That shapes how you see risk, opportunity, and time.

The years between academia and now included startup work at Decooda and Mesa Associates, circuit breaker analysis and Oracle Cloud Infrastructure. They also included periods of recalibration—understanding when to optimize for external output versus internal clarity.

Since April 2023, I've been working as an independent AI research engineer. I write technical content that occasionally reaches unexpected places (200K views on Reddit, my Medium essay shared by Jeff Dean). Currently exploring LLM evaluation, robustness, and the ways models fail across linguistic contexts—the gap between benchmark performance and real-world reliability. And I'm looking for a team where exceptional technical work is the strategy, not just the output.

Geolinguistic Analysis via Twitter

NeurIPS 2018 Workshop (Black in AI)

Kernel methods for mining dialectal variation in Lyon French. Under Jacob Eisenstein.

Stylistic Variation in Social Media Part-of-Speech Tagging

NAACL 2018 Workshop

Social network structure correlates with POS tagger errors on Twitter. Under Jacob Eisenstein.

Arabic Speech Recognition

MIT CSAIL 2015 · James Glass

Lexical modeling for Egyptian Arabic ASR. Grapheme lexicon outperformed diacritized approaches.

Featured Projects

Speech-Enabled Local Chatbot

2025

A fully local AI chatbot combining OpenAI Whisper for speech recognition with Ollama's Mistral 7B. Users speak questions, get AI-generated responses—all running on consumer hardware without external API calls.

Python Whisper Ollama Gradio FastAPI

Real-time audio transcription with immediate feedback
Hardware-optimized with configurable timeouts
Web interface via Gradio + ASGI deployment
Supports multiple audio formats and Ollama models

Disease Outbreak Detection via NLP

2020 → 2025

Manning Publications liveProject demonstrating NLP evolution in public health surveillance. Extracts diseases and locations from news headlines, performs spatial clustering, generates interactive outbreak maps. Compares 2020 traditional methods with 2025 LLM-enhanced approaches.

Python spaCy scikit-learn DBSCAN Folium

84% location extraction, 77% disease extraction accuracy
Identified 13 spatial outbreak clusters from 200 headlines
Interactive geospatial visualizations with Folium
Comparative analysis: template-based vs LLM synthetic data

Selected Work

Technical Writing Medium Reddit (#1 on r/AI)

What 5,000 Hours of Mastering Tekken Taught Me About AI Research

Autophenomenological analysis exploring parallels between competitive gaming mastery and pattern recognition in ML research. Examining whether the obsessive optimization that makes someone elite at fighting games relates to what makes someone effective at research.

~200K views · #1 on r/ArtificialIntelligence

Essay Medium

From Sudan to Silicon Valley: Beyond the Resume

Personal narrative on the gap between credentials and lived experience. What it means to be celebrated in one context and invisible in another. Shared by Jeff Dean.

200K views · Shared by Jeff Dean

These are just highlights. I've written 11 in-depth articles on AI research, personal challenges, and technical work.

View All My Writing →

Competitive Optimization

5,000 hours of competitive mastery.

Over five years, I reached the top 0.5% globally in Tekken (Tekken God rank, Nina Williams main) and held the #1 position worldwide for ranked wins. This required frame-perfect execution, matchup-specific adaptation, and systematic optimization under pressure.

The pattern recognition that makes someone elite at fighting games—reading opponent tendencies, exploiting micro-advantages, iterating on failure in real-time—is the same cognitive process that drives effective research and engineering. The difference is where you apply it.

What competitive gaming taught me: how to diagnose failure systematically, how to stay sharp under uncertainty, and when optimization for its own sake becomes the problem instead of the solution.

0.5%

Global Rank (Tekken God)

Worldwide Ranked Wins

5,000

Hours of Play

Nina

Main Character

What People Say

Jeffrey Lai Co-founder & CEO, Tallgeese AI · Ex-Apple

Taha made a lasting impact during his Siri internship, advancing Siri's i18n and scaling efforts. He built a powerful bug logging tool that streamlined debugging 10x for engineers and program managers, still in use today. His project was selected for Apple's internal research summit.

Abdelrahman Karrar PhD, IEEE Senior Member · Professor, UT Chattanooga

I would meet him at the University of Tennessee at Chattanooga when he would come on some weekends. I was struck by how different he was compared to his peers. He was interested in intellectual discussions, exploring social, cultural and political issues. He offered to accompany me on a journey to Atlanta Airport to pick up a mutual friend, and the two-hour journey turned out to be an education for me. I am still influenced by the ideas we discussed and the podcasts he made me listen to.

Jonathan Zhang Machine Learning Engineer, Pinterest

He developed a testing tool that received recognition from Apple's head of engineering and, notably, remains in use today—an uncommon accomplishment in an industry where software often changes rapidly. This achievement highlights Taha's engineering precision and foresight in building solutions with lasting impact.

Robert Citorik Co-Founder, Tessera Therapeutics · MIT

He was easily one of the most motivated students I have had the pleasure of advising, always providing questions that drove our weekly group meetings toward useful discussions, and one of the few who would connect offline for conversations regarding science and careers. We even continued these discussions after the end of the program.

Shuang Z. Tu Professor, Jackson State University

Taha was undoubtedly the best undergraduate student I had the privilege of teaching during his time at Jackson State University from 2013 to 2017. His performance in all the courses I taught was exceptional. I learned that Taha completed the Machine Learning course offered by Stanford as an independent study under the guidance of the ECE Department Chair, and once again, his performance was exemplary. I am therefore very confident that Taha is academically well-prepared for any technical position he chooses.

I bridge computational linguistics research and production systems that ship—and last.

Download Resume

Local-First Voice Agent: High-Latency Orchestration on Legacy Hardware

Why This Matters

Evolution of NLP Architectures: Benchmarking the 5-Year Collapse of Static Embeddings

Why This Matters

From Khartoum to Georgia Tech to wherever this goes next.

Geolinguistic Analysis via Twitter

Stylistic Variation in Social Media Part-of-Speech Tagging

Arabic Speech Recognition

Speech-Enabled Local Chatbot

Disease Outbreak Detection via NLP

What 5,000 Hours of Mastering Tekken Taught Me About AI Research

From Sudan to Silicon Valley: Beyond the Resume

5,000 hours of competitive mastery.

Let's Talk