I build production AI systems — the kind that survive real users, not just demo day.
5+ years shipping software, 290 public repos, 4,726 contributions and counting. I work across the whole stack: training embedding models from scratch, wiring LLM agents into real workflows, and building the front-ends and APIs that make it usable.
By day I run MenteE — the AI product studio where the recent work happens (the products speak for themselves below). My own site lives at syab.tech.
- 🔭 Now: multilingual embedding research, agentic tooling, and hiring/workflow automation
- 🌍 Based in Riyadh, Saudi Arabia · working Pakistan → Gulf → global
- 🤝 Open to: AI product builds, full-stack engineering, technical partnerships, startup collaboration
🧮 How the Level 5 and language ranks are calculated (all numbers come from the GitHub API — nothing made up)
Level formula
score = contributions + commits + (5 × PRs) + (5 × stars) + (2 × repos)
= 4,726 + 3,824 + (5 × 123) + (5 × 26) + (2 × 290)
= 9,875 → Level 5 of 6
L1 < 1,000 · L2 1k–3k · L3 3k–6k · L4 6k–9k · L5 9k–14k · L6 14k+
Languages by repos authored (222 own repos): Python 72 · JavaScript 36 · Java 32 · C++ 15 · HTML 12 · TypeScript 11
Other API facts: GitHub stats rank B+ · contributed to 13 repos in the last year · 68 forks · member since Mar 23, 2021 · 59 followers · 142 following
Day-to-day work at MenteE — products, not pitch decks.
AI hiring & interview platform. AI job matching for candidates, ATS-optimized CV scoring, unlimited AI mock interviews, and full applied → hired pipeline automation for organizations. Built for Pakistan, expanding across the Gulf (UAE · Saudi Arabia · Qatar · Kuwait · Oman · Bahrain).
Flask · PostgreSQL · Redis · Kafka · LLMs
Platform · Live app
A terminal-native software-engineering agent: it investigates your repo, plans the change, edits it, runs your tests and reports back with evidence. Model-agnostic, MIT licensed, zero-install try via npx.
Agents · CLI · Any model provider
Details
Compact multilingual text embedding models trained from scratch — no borrowed weights: ~41M parameters, 384-dimensional output, Arabic + English + Urdu retrieval, two-stage MLM → contrastive recipe, Apache 2.0.
PyTorch · Embeddings · Single-GPU training
Research · Models · Preprint (DOI)
mentee-embed — competitive multilingual embeddings, trained from zero.
- 🧠 ~41M params · 384 dims · 3 languages (العربية · English · اردو) — low-resource languages treated as first-class
- 🥇 Two-stage from-scratch training: masked language modeling, then contrastive fine-tuning — a cheap bootstrap that avoids representation collapse
- 📄 Preprint: mentee-embed-v3 · Zenodo · 26 Aug 2026 · CC BY 4.0 → DOI 10.5281/zenodo.22117673
- 📊 5 publications, measured numbers instead of marketing — failures documented next to wins, every artifact public
~50 projects live on syab.tech — here are the ones worth your time.
🤖 AI, agents & LLMs
- Req2Ops
LangGraph · FastAPI · React 19— turns messy briefs (text, PDFs, audio, screenshots) into structured project plans: 56 plan types, sprint estimation, risk scoring, live SSE streaming. - Mockster.ai
Python · LLMs— AI interview simulator for realistic, repeatable mock interviews. - AI Voice Agent
FastAPI · Retell AI— production voice agent for automated appointment booking with a live sheet-based CRM, email and webhooks. - DOCX AI Translator
Python · Groq Llama 4 · python-docx— Chinese → English document pipeline: 668/668 blocks translated on a 55 MB paper, images preserved, checkpoint/resume, quota-aware retries. - University Chatbot
Python · NLP— campus assistant answering from real university knowledge. - PodcastMaker
OpenAI · Streamlit— text to podcast: LLM scripting plus TTS audio generation.
🧰 Platforms, apps & tooling
- StitchPoint
React · Node · MongoDB · Stripe— multi-vendor artisan marketplace with admin dashboards, vendor analytics, orders and secure payments. - RunAI
React · JSON-driven— gamified AI/ML learning: mission-based challenges, embedded editor, terminal simulator, XP progression. - Menteemart
React · Node— full commerce build. - DocsBox
Kotlin · Firebase · AES— privacy-first encrypted document storage. - CamWatch
OpenCV · YOLO— vision-based monitoring pipeline.
📦 Open source on PyPI
- datafit — data preprocessing + feature engineering in a couple of lines.
- menteedb — minimal, typed Python database layer.
- autodatap — automated ML data preparation toolkit.
🌐 Client work — shipped & live
Nainzaka Aesthetics · Bombay Town · Arabian Deccan · Mashwi Singary · Machali Kingdom Riyadh · Sabqa · Laboratory Management System · Global Craft Hub
🎓 Learning tracks I built for other people (free, open source)
Notes from the product and research trenches, on the MenteE blog:
- 📌 Jobs in Dubai, UAE and the Gulf from Pakistan — how AI matching actually finds roles you'll get · 8 min
- 🎤 AI mock interviews that prepare you for real interviews in Pakistan and the Gulf · 7 min
- 🏢 Hiring in Pakistan and the Gulf: pipeline automation that makes it faster and fairer · 7 min
Core expertise: LLMs RAG Agents Embeddings Fine-tuning Computer Vision NLP FastAPI PostgreSQL System Design MLOps CI/CD
- 🥇 Founded MenteE — an AI product studio shipping products, not services (menteeai.org)
- 🧠 Trained embedding models from scratch (~41M params) on a single GPU and published the research openly (Zenodo, CC BY 4.0)
- 📦 3 packages on PyPI —
datafit·menteedb·autodatap - 📈 4,726 contributions in 5 years,
B+GitHub rank, longest streak 109 days - 🌍 Products live across Pakistan & the Gulf — with Arabic and Urdu support rolling out
- 🎓 Mentored through open learning tracks used by students and junior devs
Agentic AI · Advanced RAG · Vision Transformers · Representation collapse · Production MLOps · Multilingual retrieval · Scalable AI infra
If you've got an AI product, a messy workflow, or a system that needs to actually scale — I'm interested.




