Danish Laraib // AI Systems Engineer.

Enterprise supply chain networks struggled with volatile procurement lead times versus inter-facility warehouse transfer trade-offs. Decisions made by disparate inventory and logistics teams lacked explainability, risking stockouts or excessive transit costs.

Engineered a Domain-Driven Typed Property Graph (TPG) on Neo4j 5.x unifying 12 intelligence layers into a causal market simulator. The decision engine evaluates trade-offs between procurement lead times and transfer surcharges, producing mathematically sound, auditable recommendations with unified multi-role dashboards.

Customer support voice and chat infrastructure handled 1,500 daily interactions where over 30% contained multiple concurrent intents (e.g. broadband failure coupled with tariff downgrade requests). Traditional flat routers suffered from prompt token bloat, high latency, and frequent misrouting into general support queues.

Constructed a stateful LangGraph orchestrator that embeds discrete agent capability vectors for routing without dumping lengthy instruction sets into LLM context windows. Incorporated a persistent memory harness to retain multi-turn context, an automated recovery node to intercept incorrect classifications, and quantized RAG chunking to accelerate retrieval passes.

50K+ unstructured open-domain customer interactions contained intent gaps between automated bot flows and human escalation queues. Categorizing them manually was intractable, while flat classifiers failed to capture nuanced hierarchical parent-child relationships.

Implemented the TnT-LLM methodology (Wan et al., Microsoft Research) to turn unstructured questions into a reusable, readable hierarchical topic map. Utilized UMAP dimensionality reduction, HDBSCAN clustering, and LoRA-finetuned LLM label proposal to stamp every interaction with verifiable routing tags for batch and near-realtime dispatch.

Deployment across 10 international markets mandated strict regulatory privacy compliance (GDPR and national data protection standards). Customer dialogues contained IBANs, phone numbers, and identifying records that had to be sanitized with sub-millisecond overhead before ingestion into LLMs.

Built a hybrid PII anonymization pipeline integrating Microsoft Presidio, language-tuned spaCy models, and compiled deterministic regex recognizers. Evaluated four enterprise guardrail frameworks on real interaction data and productionized a fine-tuned Phi-4 safety filter achieving 99%+ accuracy across 25+ entity types.

A high-traffic FAQ conversational agent serving 200,000+ monthly user interactions suffered from silent retrieval degradations, periodic chunk truncation in document scrapers, and unmonitored embedding latency drift across multilingual knowledge bases.

Engineered an automated continuous evaluation harness computing precision@k, recall@k, and hallucination metrics on synthetic and real query splits. Reproduced historical failure cases, diagnosed and resolved a silent chunk-truncation bug in the scraper ingestion pipeline, and benchmarked multilingual embedding alternatives for throughput and quality.

Launching conversational agents into new country markets previously required building, training, and validating separate language models per locale, driving up training costs, cloud memory footprints, and ongoing maintenance complexity.

Engineered a unified multilingual routing classifier trained on 150K conversational samples. Combined fine-tuned XLM-RoBERTa with LoRA-adapted Gemma3-27B to produce a single generalized model achieving 99%+ routing precision across languages without requiring country-specific model retraining.