Representative production work. Client-sensitive implementation details are intentionally abstracted.
01
Recommendation intelligence
Semantic discovery at streaming scale
A RAG-based content discovery architecture pairing vector retrieval with tuned generation for more relevant, explainable recommendations.
DECISION NOTESelected RAG over end-to-end fine-tuning to balance retrieval quality, operating cost, and the pace of catalog change.
+30%discovery accuracy
RAGPineconeFAISSFastAPIKubernetes
02
NLP · financial systems
Real-time fraud intelligence
A production fraud detection system combining BERT language signals with XGBoost decisioning for high-volume banking workflows.
DECISION NOTETurned unstructured transaction context into actionable risk signals while preserving a path for review and iteration.
$2M+annual savings
BERTXGBoostPythonAWSMLOps
03
Applied language AI
Conversational support engine
A transformer-powered support experience built to resolve common requests quickly and hand off complex conversations cleanly.
DECISION NOTEDesigned around resolution—not novelty—with intent routing, reliable fallbacks, and measurable customer outcomes.
50K+daily interactions
RasaTransformersNLPREST APIsMonitoring
BASED IN THE U.S. · WORKING GLOBALLY
02ABOUT
I care about the space between a promising model and a product people can trust.
I’m an AI/LLM engineer with 5+ years across intelligent systems, scalable backends, and full-stack products. My work spans model selection, retrieval design, inference optimization, and production operations.
I’m most useful where the problem is still a little messy: when accuracy, latency, cost, and user experience all need a seat at the same table.
