Work

Problem Software defects that are not flagged by the compiler, the static analyzer or a code review still surface during vehicle integration testing, which is the most expensive place to find them.

Built A multi-agent repo scanner on Amazon Bedrock AgentCore that reasons across the codebase to identify software issues.

Outcome Identified more than 30 high-priority findings that earlier tooling and reviews had missed, each triaged by an automotive domain expert.

Problem Undocumented mainframe code stalls code understanding, feature development and enterprise migration. The people who wrote it have moved on, so the knowledge is very scarce.

Built Agentic pipelines on Amazon Bedrock that generate documentation from legacy sources, across 1.3 million lines of NCL.

Outcome Generated documentation reflecting business logic at three levels of detail targeted at different audiences, from migration engineers to business summaries. The project has moved to production rollout.

Problem Copywriters needed 2-3 days per campaign to localize marketing content for each major market. The task is not only translation but market tonality, regulatory terms and brand voice.

Built Localization pipeline with a self-reflection layer, where the reflection depth and compute budget is chosen by benchmarking 10 models across four domains against the quality, cost and latency frontier.

Outcome On German, the share of localizations rated at least as good as a copywriter's rose from 38% to 47%. French and Spanish, already at 61% and 49%, gained little, which is what made per-market evaluation crucial.

Problem RAG and generic coding assistants were not enough on a large automotive codebase, where conformance to industry-specific rules and patterns is important.

Built A multi-model coding assistant pairing Qwen3-32B fine-tuned on automotive production code with an orchestrator powered by Claude Sonnet and verification tools that let the agent check the coding patterns.

Outcome The fine-tuned model, trained on roughly 7,000 annotated functions, cut average Automotive Architecture rule violations from 1.22 to 0.02 against the prompt-engineered baseline.

Problem Maritime analysts drown in isolated alerts and anomalies that carry no context on their own. Most of the work is collecting data rather than deciding anything.

Built An agentic pipeline that fans out to news, web search and weather sources, powered by Amazon Bedrock LLMs judging whether what came back explains the event and writing summary explanations.

Outcome Shipped as part of Windward's MAI Expert, turning the hours an analyst spent gathering context per alert into a cited report.

Problem A library of medical articles goes stale as clinical guidance moves, and re-reviewing it by hand does not scale with the library.

Built MACROS, an AI-powered system on Amazon Bedrock orchestrated with AWS Lambda and Step Functions, which reviews articles against guidelines and proposes revisions.

Outcome The proof of concept reached 80% accuracy and over 90% recall, cutting per-guideline review from hours to minutes.

Problem Extracting structured data from PDFs and emails at scale, where one managed extraction path does not suit every regulatory and regional constraint.

Built An end-to-end AWS CDK application on Amazon Bedrock Data Automation, Step Functions and Lambda, with three interchangeable extraction paths: managed BDA, custom Bedrock models, and Textract OCR paired with an LLM.

Outcome Published under MIT and deployable as a single CDK stack, so teams adopt the pipeline rather than rebuild it.

Problem Multi-source research is the same shape of work in every field, and an agent left to search first wanders, writing thin sections it never notices.

Built An agent on Amazon Bedrock AgentCore that scaffolds an empty report skeleton before searching, runs four to six searches judging source quality, writes section by section with inline citations, then re-reads itself and fills the gaps.

Outcome Scores 24.0 on HLE-Search and 49.6 on GAIA, ahead of Grok DeeperSearch on both benchmarks and of Perplexity Deep Research on HLE-Search. Ships with over ten connectors and deploys as a single CDK stack.