Abdullah Naeem, AI/ML engineer in Lahore. I build AI that listens.

Selected work

Five systems, three of them in production. Each write-up covers the problem, the data, the architecture, the results, and where it breaks.

1. Arthus Systems, 2026

   AI phone intake for US law firms

   Answers intake calls, screens callers for conflicts of interest mid-call, and refuses to give legal advice. Compliance steps run in code, not in the prompt.

   • <800 ms reply time
   • <1 s conflict check
   • 52 adversarial prompts in CI

2. Teaching Circle, 2026

   WhatsApp enquiry automation

   Answers student questions about courses, payments and enrolment on WhatsApp, replacing a support queue that took two to three people.

   • 100+ enquiries a day
   • 95%+ classification accuracy
   • 80% less manual work

3. Final-year project, 2026

   Driver drowsiness detection

   Watches a driver through an ordinary webcam on a CPU-only laptop and raises an alert when eyes stay closed or yawns pile up.

   • 99% eye-state accuracy
   • 96% yawn accuracy
   • 40+ FPS on CPU

4. Group project, 2026

   PsychConnect telepsychology platform

   Gemini drafts assessments of drawing tests and PHQ-9 and DASS-21 questionnaires for licensed psychologists to review, never to diagnose.

   • 20–30% faster assessments
   • 3-phase Gemini pipeline
   • 9 clinical safety rules

Earlier and smaller projects

• Housing price prediction Compared five regression models on Taiwanese housing data; polynomial regression won with R² ≈ 0.73.
• VGG-16 from scratch and YOLOv8 Built VGG-16 in PyTorch for CIFAR-100, fine-tuned YOLOv8-L, and wrote non-maximum suppression by hand.
• Movies database Normalised MySQL schema with 9,500 entries and analytical queries, built as a group project.
• Space Invaders in Pygame A clone with collision detection and scoring, from my first programming course.

Experience

1. Sep 2026 – now

   AI/ML Engineer, Miras AI

   Lahore

   • Adding Urdu to NVIDIA Nemotron 3.5 for low-resource voice agents: extended the tokenizer with a new Urdu token and fine-tuned on 121 hours of speech on a vGPU T40.
   • In testing so far: normalised WER of 22.1% and CER of 17.0%, down from an effective 100%, keeping 74.6% of English words in code-switched speech.
   • Streaming tests at 0.56 s algorithmic latency and 67 ms p95 compute per chunk, a real-time factor of 0.18. Still improving week by week.

2. Jun – Aug 2026

   AI/ML Engineer, Arthus Systems

   Remote

   • Built a real-time phone agent for US law-firm intake in TypeScript and Fastify, over a persistent WebSocket to Retell AI and the Claude API, replying in under 800 ms.
   • Enforced each compliance step with a state machine and screened callers for conflicts mid-call in under a second with PostgreSQL trigram matching.
   • Stopped unlicensed legal advice with prompt rules plus an output filter, tested against 52 adversarial prompts in CI on every change.

3. Apr – Jul 2026

   Software Engineer, AI Automation (contract), Teaching Circle

   Remote

   • Designed and deployed a WhatsApp assistant routing 100+ student enquiries a day through GPT-4 and GPT-5.1 mini, orchestrated with LangChain.
   • Built a webhook pipeline from student form submissions, classifying enquiries at 95%+ accuracy across 100+ automated tests, with PostgreSQL logging.
   • Cut manual response work by 80%, replacing a two-to-three-person support team; containerised with Docker, with no recorded outages.

4. Graduated Jun 2026

   BSc Computer Science, Forman Christian College University

   Lahore

   Machine learning, deep learning, database systems, data structures, software engineering, web development.

Tools, and where I’ve used them

Every tool here is tied to real work; each dot is a project that used it. Select a project to light up its stack.

Certifications

Courses I’ve finished. Each one links to the issuer’s own page, so you can check it in a click.

Writing