Senior Scientist at AWS FDE. I publish applied AI research and embed with engineering teams to build custom AI systems, from experimentation to production deployment.
8+ years in AI/ML · Focused on LLMs and agentic systems
BlogBuilder Center
Language models solved only 34% of SWE-InfraBench's infrastructure-as-code tasks in one attempt. We ran Kiro over all 100 and found that a full agent loop with Sonnet 5 reaches 82%, with partially correct answers converging once the agent can re-run the tests itself.
View all 29 blog posts
TalkAWS Summit · Zurich
Learn to secure LLM applications through red teaming. We will explore prompt injection, jailbreaking, toxic output generation, and secure code execution risks in AI agents. See live attack demos and hands-on threat modeling. Discover how Amazon Bedrock Guardrails, Amazon Bedrock AgentCore, and monitoring tools can help defend against vulnerabilities. Leave with actionable strategies for hardening generative AI systems and conducting security assessments.
View all 36 talks
PaperCOLM 2026 Workshop on Tokenization
A tokenizer defines the vocabulary a language model scores at every decoding step. Output-vocabulary pruning modifies this vocabulary after training: it keeps the model and tokenizer fixed and restricts the set of output tokens scored at inference, compressing the vocabulary-sized output layer. The dominant approach, frequency-based pruning, optimizes average token coverage, which may lead to stripping the tokens that minority languages and specialized domains depend on. We address this by giving the selection problem an explicit objective: the expected log retained probability mass under the deployment distribution, which we show is equivalent to minimizing reverse KL to the teacher and is monotone submodular. We therefore select the retained tokens by greedy maximization of this objective, which inherits a (1 - 1/e)-approximation guarantee. Frequency pruning is its linear surrogate, optimal only when contexts agree on a common token ranking. Across synthetic and multilingual deployments, greedy outperforms both frequency- and magnitude-based pruning, with the largest gains where that agreement breaks down. Under English-dominant calibration, greedy cuts Arabic reverse KL by 0.53 nats and improves Russian generation by +1.4 chrF, at negligible cost to English; the same support raises speculative-decoding draft acceptance by +5.4pp on Arabic. Our main experiments use Qwen3.5-0.8B, and a cross-family check on Gemma-3-1B confirms the qualitative pattern.
View all 18 papers
Agentic code scanning system that found 30+ high-priority defects in a production automotive codebase.
Agentic pipelines that recovered the business logic buried in 1.3 million lines of mainframe code, now moving to production rollout.
Test-time self-reflection for content localization across 17 European markets, tuned for quality, cost and latency.
Open to collaboration, speaking invitations, and conversations about applied AI.