I build AI products people actually use.

ai · ml

• Built the company's first production LLM inference stack: a FastAPI and vLLM service handling concurrent token generation, serving several hundred concurrent requests at peak.
• Designed the RAG retrieval layer on Qdrant over several hundred thousand indexed documents, using asynchronous embedding and vector search to keep retrieval off the critical path so lookup cost does not compound with generation latency under load.
• Built complexity-based query routing between self-hosted open-weight models and OpenAI GPT-4o, cutting per-query API spend and establishing the evidence base for the platform's model-sourcing decision.

frontend

• Architected the real-time streaming interface: a React client consuming the inference API over SignalR, with token-by-token rendering, scroll anchoring, request cancellation, and recovery from partial or failed streams.
• Built the component and state layer for the RAG experience, covering citation surfacing, source previews, and skeleton loading, using memoization and Suspense boundaries to stay responsive during multi-second retrieval and generation.
• Established shared frontend conventions: typed API contracts against the FastAPI backend, component composition rules, and error boundaries, removing duplicated integration code across every AI feature.

both

• Established the concurrency, backpressure, and error-recovery patterns now standard across every AI feature in the product.
• Set engineering standards for AI feature delivery covering prompt versioning, model evaluation before release, and observability on inference latency and token spend.

ai · ml

• Shipped a retrieval-augmented assistant for [client or domain], indexing [N] documents into [Qdrant or Pinecone] and cutting [manual lookup time] by [X%].
• Integrated [Gemini or GPT-4o] into [product], owning the prompt and evaluation pipeline and lifting [answer accuracy or task completion] by [X%] against a labeled benchmark.

frontend

• Built the streaming chat interface in Next.js and TypeScript with token-by-token rendering, request cancellation, and recovery from failed streams, bringing time to first token down to [X ms].
• Rebuilt [client]'s [surface] in React and Next.js, raising Lighthouse performance from [X] to [Y] and cutting bundle size [Z%].

both

• Took [N] of the four engagements from prototype to production deployment on [AWS or GCP] with Docker and CI/CD, handing each off with documentation the client's own team could build on.

ai · ml

• Led LLM integration across 10+ client products, delivering structured extraction, summarization, and conversational agents on shared internal patterns that cut AI integration time on new projects from weeks to days.
• Engineered prompt pipelines and context-injection strategies against the Gemini API, evaluated against human-labeled benchmarks rather than spot checks, increasing user engagement 25%.
• Redesigned data retrieval on ML-serving paths through query restructuring, indexing, and a caching layer, reducing query latency 40% and removing the dominant bottleneck in end-to-end inference response time.

frontend

• Led frontend development for 10+ client products in Next.js and React, spanning data-heavy dashboards, multi-tenant platforms, and AI-driven tools.
• Built real-time conversational AI interfaces with the Gemini API and Vercel AI SDK, covering streaming responses, structured output rendering, and conversational state.
• Improved page performance through code splitting, image and font optimization, and server-component adoption, tightening Core Web Vitals on client-facing pages.
• Established component standards with product and design partners, holding the line on semantic markup, keyboard navigation, and WCAG 2.1 AA contrast requirements.

both

• Owned deployment on AWS with Docker and GitHub Actions CI/CD, automating Jest and Cypress runs, type checks, linting, and preview deployments across a distributed team.

ai · ml

• Led R&D for hate-speech and misinformation detection in Amharic, Afaan Oromo, and Tigrinya, owning the full lifecycle from dataset construction through production deployment.
• Fine-tuned multilingual BERT on a purpose-built labeled corpus, applying transfer learning to compensate for severe data scarcity and class imbalance across all three languages.
• Designed the labeling taxonomy and annotation guidelines that made consistent multi-annotator training data possible at scale, the foundation the entire modeling effort depended on.
• Deployed automated content-analysis data pipelines on Google Cloud AI, cutting manual review time 35% and scaling verification throughput well beyond human moderation capacity.

frontend

• Built and owned the complete React frontend for the internal ML analysis platform, giving researchers interactive visualization of model outputs over large, noisy datasets with D3.js and Recharts.
• Designed data-dense interfaces with filtering, drill-down, and real-time result streaming against the Python and TensorFlow backend, making complex model behavior legible to non-engineers who had to act on it.

both

• Used engagement analytics and A/B testing to direct training data selection and public-facing component decisions, contributing to a 170% increase in traffic over six months.

frontend

• Launched a cross-platform Flutter app for campus navigation and real-time announcements, reaching 2,000+ active users in its first year.
• Built the full frontend for a university procurement application, designing a flow that cut manual processing effort 65%; still in daily use across the university.

both

• Conducted requirements research with 250 university students and translated the findings into the prioritized specification and information architecture for both products.