AI Engineer — LLM applications, RAG systems and agentic automation
Production over notebooks.
I build AI systems that run in production, not in notebooks. Currently working independently with EU and US clients on retrieval pipelines, multi-agent automation and the FastAPI services that hold them together. Previously led a small AI team at Tkrupt.
A 0.5B language model that repairs Arabic OCR output — the step between "we scanned the archive" and "the archive is searchable."
No paired Arabic OCR corpus exists, so the supervision is generated: a script-aware confusion model that breaks clean text the way a scanner actually breaks it — dot and skeleton confusion, word splits and merges — weighted by real failure modes. A production guardrail rejects rewrites rather than repairs.
| CER | WER | |
|---|---|---|
| Raw OCR (do nothing) | 0.0808 | 0.4112 |
| Untuned Qwen2.5-0.5B | 1.8123 | 2.2680 |
| Fine-tuned + guardrail | 0.0671 | 0.2042 |
Word error 41.1% → 20.4% — a 50.3% reduction. Trained in 80 minutes on a single laptop GPU. Runs on CPU, on-premise.
LoRA PEFT Qwen Arabic NLP Small Language Models
Retrieval-augmented assistant over Pinecone — document ingestion, embedding and semantic search with an LLM answer layer.
RAG Pinecone Vector Search Embeddings
| 🔻 10–20x | cost-per-output reduction on a live LLM pipeline (~$1K/mo → $50–100/mo) |
| 📚 5,000+ | document corpora in production RAG pipelines — analyst review cycles cut 50% |
| ⚡ Sub-5s | responses at 50+ concurrent users, serving 1,000+ users |
| 🔌 40% | reduction in new-integration time via a modular MCP-style tool layer |
| 🛡️ 60% | reduction in production pipeline failure rates |
| 👥 3 | engineers led at Tkrupt |
|
|
|
|
Languages
AI / ML
LLM Providers
Backend & Data
Vector Stores
Infrastructure & Automation
| Role | Where | When |
|---|---|---|
| Independent AI Consultant | Remote — EU & US clients | Aug 2026 – present |
| AI Integrations Engineer | Tkrupt, Lahore | Apr 2025 – Aug 2026 |
| AI Engineer | Zaltech AI, Lahore | Jul 2024 – Mar 2025 |
| Freelance Full-Stack Developer | Fiverr | Feb 2022 – Aug 2024 |
Co-founder. Deep-learning image steganography for digital rights protection and content-ownership verification.
🌍 Exhibited at GISEC Global 2025, Dubai World Trade Centre 🥈 Runner-up, Zindigi Prize 2024 Islamabad Regionals — against 85+ startups 🚀 Selected for National Incubation Center Lahore, Cohort 3
const exploring = {
protocols: ["MCP server design", "agent interoperability"],
evaluation: ["LLM-as-judge harnesses", "retrieval eval at scale"],
serving: ["quantisation & model compression", "on-prem CPU inference"],
research: ["Arabic NLP", "synthetic supervision for low-resource tasks"],
};BS Computer Science — Pakistan Institute of Engineering & Applied Sciences (PIEAS), 2020–2024 ISC2 Certified in Cybersecurity (CC) · LangChain for LLM Application Development (DeepLearning.AI)
Open to AI engineering roles — remote, Lahore or GCC.
musabjaved47@gmail.com