Work

Restaurant and hospitality managers spend their day sorting: timecards, time-off requests, schedules and coverage gaps. The platform held the data, but the work of reading it and acting on it was still manual.

Led the GenAI team that took Salli from prototype to General Availability. Salli is a multi-agent system on LangGraph: an orchestrator with fan-out routing hands work to specialist agents, each running its own ReAct tool-calling loop with per-agent model selection, token streaming and a visible reasoning stream. Langfuse handles prompt management, tracing and usage analytics.

General Availability in April 2026, with 10+ production agents covering scheduling and labor, timekeeping, talent acquisition, CoreHR, compliance, local-context awareness and multi-location management. The Local News & Events agent feeds what happens around each restaurant into demand forecasting.

Salli started as read-only Q&A. Managers wanted it to do the work, but an agent that changes schedules or timecards has to be safe to trust.

As Feature Lead for the Agentic Actions program, moved Salli to executing workforce-management actions. Every action sits behind a human-in-the-loop confirmation gate: Salli recommends, the manager makes the final call. A dedicated regression suite covers the action set.

80 executable workforce-management actions, each requiring human confirmation.

Shipping agents one at a time does not scale. Each new agent meant re-deciding architecture, tool boundaries and tests.

Built a repeatable workflow: architecture discovery, a human validation gate, then scaffolding of agents, tools, graph nodes and tests, ending in a spec document and a PR handoff. It codifies conventions such as service-layer tool boundaries and a cap on tools per agent to keep routing reliable.

Architecture documentation is generated as part of the workflow.

LLM applications are only as good as the data behind them, and that data keeps changing in operational databases.

Built RAG applications, knowledge graphs, chatbots and LLM-based automated components such as auto-mappers and classifiers, integrated with vector databases. Prototyped Conduit pipelines that stream Postgres changes (CDC) into Weaviate, so a vector index stays in sync with the source database.

Bank employees needed fast, accurate answers on Arabic-language policies and procedures, and help handling customer inquiries.

Designed and built an agentic, RAG-based conversational assistant. Architected the retrieval and multilingual prompt-management layer, and built evaluation pipelines that continuously validate response accuracy and policy compliance before rollout.

Reviewing patient calls and provider performance by hand was slow and hard to keep consistent.

Designed the pipeline for an AI Service Coordinator Analyst that audits patient calls through transcription, summarization, structured-insight extraction and progress tracking. Built an AI Supervision & Auditing agent that scores provider performance against configurable templates with conversational reporting, plus a company-wide employee Q&A knowledge-base agent. Delivered the evaluation pipelines, analytics dashboards and reporting tools behind them, with guardrails to keep automated scoring consistent and defensible.

Beginner learners need a tutor that holds a coherent conversation at their level, not one that drifts into language they cannot follow.

Architected the end-to-end pipeline for an AI avatar classroom assistant, using context engineering and iterative prompt optimization to keep multi-turn tutoring conversations coherent and level-appropriate.

Restaurant and hospitality operators schedule staff against expected demand. If the forecast for sales and other demand metrics is off, the schedule is off: too many people on a quiet shift, too few on a busy one.

I built my time-series forecasting experience earlier in my career, including demand-prediction work at Foothill Technology Solutions. At Harri I now manage the Data Science & ML team and give technical supervision and support on the AI Demand Forecasting service, which forecasts sales and other metrics. On the agentic side, Salli's Local News & Events agent captures what happens around each restaurant and feeds it into demand forecasting.

Harri customers rely on the forecasts for labor deployment and scheduling.

Endometriosis is often diagnosed years after symptoms begin, and there is no simple test for it. A femtech client wanted to know whether patterns in patients' medical histories could flag it earlier.

Led a team of 3 data scientists across 3 client engagements in healthcare and social media. For this one, we worked with US healthcare data, using ICD diagnosis codes to describe patient histories and OHDSI ATLAS to define and analyse patient cohorts. On that basis we built predictive early-diagnosis models. Other engagements used NLP and embedding-based language models for topic modelling and clustering of large unstructured text datasets.

The early-diagnosis models were used in the client's research publication.

Harri's workforce management customers in hospitality needed ML-powered services that work on their own operational data.

Owned the full lifecycle of each service: data wrangling, feature engineering, training with classical ML and deep learning, validation and deployment. Built MLOps pipelines on AWS SageMaker that automated data processing and model deployment, and did R&D on personalizing the services for the hospitality vertical. Started as an intern building an NLP resume parser for Harri's hiring workflows.

3 ML services in production for Harri's workforce management customers. The resume parser was integrated into Harri's hiring and screening workflows.

Conduit, Meroxa's open-source data streaming platform, connects systems through connectors written in Go against the Conduit connector SDK. Every new source or destination meant writing another connector by hand.

Worked on automated code generation and execution with LLMs: fine-tuned GPT models on the Conduit connector SDK and used them to generate connector code into Conduit's standard Go connector template, covering the source, destination, configuration and tests.

Organizations running on legacy Oracle Forms applications and scattered data want to modernize their systems and put AI to work on top of them.

Consulting on agentic engineering, building data pipelines, and migrating Oracle Forms applications.