Abdul Haque
Applied AI Engineer · Production agents, LLMs and backend systems
Islamabad, Pakistanabdulhaque.dev@gmail.comlinkedin.com/in/abdul-haque-devgithub.com/Ahaque-AI
Summary
Applied AI Engineer building production LLM microservices (FastAPI) and multi-agent RAG systems (LangGraph, CrewAI) on Azure, including a speech and LLM pipeline serving 300+ sessions a day for a healthcare client. I lead agents rather than vibe code: I set the spec and own the architecture and the failure modes.
Skills
- AI and LLM
- LangGraph, LangChain, CrewAI, Multi-agent systems, OpenAI and Anthropic APIs, Azure OpenAI, Azure AI Foundry, vLLM, Ollama
- Backend
- FastAPI, REST APIs, Microservices, AsyncIO, Pydantic, OAuth and JWT, Middleware
- Data and RAG
- PostgreSQL (pgvector), Neo4j (Cypher), GraphRAG
- Cloud and infra
- Azure (AKS, DevOps, Blob), AWS (boto3), GCP, Docker
- ML, NLP and speech
- PyTorch, HuggingFace Transformers, Scikit-learn, Faster-Whisper
- Messaging and observability
- Azure Service Bus, Kafka, MLflow, Prometheus, Grafana
- Languages
- Python, SQL, Bash, C++
Experience and education
Associate Software Engineer, AI
Dec 2025 – PresentDeltaShoppe (Pvt.) Ltd, Islamabad, partner Kermit Tech (Norway)
- Sole backend engineer rotating across three concurrent client engagements in a small AI engineering pod.
- AI Therapy: first shipped Faster-Whisper and Phi-3.5-Mini (INT8) on CPU-only 15 GB AKS pods, with a multi-processing worker pool that fixed lock contention and ONNX out-of-memory failures. Since moved to Azure AI Foundry: GPT-5.4 mini and batched audio transcription, on Azure resources only in the region the client specified. JWKS token verification, encrypted audio, dead-letter queues, MLflow.
- Student Companion: propositional chunking (about 10:1 compression), tool injector pattern for per-user permission scoping, PII redaction middleware, Azure DevOps CI/CD.
- Fraud Detection: Kafka pipeline with PII masking on 6 fields and 3 Grafana dashboards with CRITICAL alerts.
Data Science Intern
Apr – Nov 2024Securely Innovations (Pvt.) Ltd
- Multi-threaded collection pipeline over about 3M automotive parts records, 3x faster than Selenium-based scrapers, by replicating the network-level API.
Computer Vision and Generative AI Intern
Jun – Aug 2024DataInsight Lab, FAST-NUCES
- Fine-tuned Stable Diffusion (DreamBooth, LoRA, QLoRA) for interior design generation on 2×T4 GPUs using model parallelism.
B.Sc. Data Science
Aug 2021 – Jun 2025FAST-NUCES, Islamabad
- CGPA 3.24/4.0, final-year GPA 3.75/4.0. Dean's List in Fall 2024 and Spring 2025. Final-year project: GraphRAG-Causal.
Projects
Synverk
20265+ active beta users. A taskboard where agents do the work through MCP and only a human can mark it Done.
AI Therapy Microservice
2025–26300+ sessions a day. Event-driven audio pipeline for a healthcare client, now on Azure AI Foundry.
AI Student Companion
2025–2690% fewer RAG tokens. Multi-tenant RAG with per-client isolation and a memory that cut 100K tokens to 10K.
5 min between regulator reports. Real-time fraud pipeline for regulated retail banking, reporting to the regulator.
GraphRAG-Causal
202582.88% F1 on CausalNewsCorpus. Graph-augmented causal reasoning in news, on par with fine-tuned BERT-Large.
IntentLens
20255,000+ reviews analysed. Four agents (scraper, intent analyser, search, supervisor) judged by an LLM.
3M records collected. Multi-threaded collection through the network-level API, 3x faster than Selenium.
2×T4 GPUs in parallel. Stable Diffusion fine-tuned with DreamBooth, LoRA and QLoRA on two GPUs.
Certifications and leadership
- Generative AI with LLMs. DeepLearning.AI and AWS
- Section Leader, Code in Place. Stanford University, 2025. 95% student retention across a global cohort.