Search is my depth. Models are my craft. Production is my responsibility.
I'm marigure. I build search, recommendation and LLM systems, owning the path from training signals to the serving contract. My edge is model depth with systems judgment: I can improve what the model learns, explain why it helps, and engineer the release.
I connect retrieval, ranking and personalization into one product decision. I train representations, choose negatives, turn behavior into learning signals, and evaluate the experience after filters and product constraints.
Contrastive learning · Neural retrieval · Transformer user models · LambdaMART
I adapt language models to specific work: extraction, grounded answers and controlled actions. I use LoRA and independent evaluation to judge the model; authorization, confirmation and recovery are explicit parts of the execution design.
LLM adaptation · RAG · Evaluation · Tool execution
I build the data and runtime around the model: CUDA / DDP training, versioned features, fast inference and compatible releases. Memory, throughput, observability and failure modes belong in the design from the start.
GPU training · Model serving · ML data · Release engineering
Prove the gain. Budget the cost. Design for failure.
I care about what users get, what it costs, and what happens when a dependency fails. The interesting work is choosing the objective, exposing the trade-off, and making the entire decision defensible.
A few receipts
- Shared retrieval: three product teams, a 160K+ item catalog, and +2.3% relative qualified watch time in user-level A/B testing of the shared release, with relevance and licensing guardrails.
- Neural inference: reranker p99 690 → 390 ms, at the same 40 RPS and traffic mix, without increased failures.
- LLM adaptation: mandatory-constraint exact match 89% → 96% on held-out dialogues, versus the prompt-only version of the same 7B model.
- Controlled actions: exact-argument confirmation, authorization, idempotency and status reconciliation. Timeout and restart tests found no duplicate operations in the tested scenarios.
Modeling · LoRA · DDP · LightGBM · CatBoost
Retrieval & inference · OpenSearch · Qdrant · ONNX Runtime · vLLM
Data & delivery · SQL · Kafka · Airflow · MLflow · AWS EKS
Based in Vietnam/Almaty · Remote work & visa-supported relocation.