AI and LLM product work
I focus on shipping useful behavior, not novelty for its own sake: retrieval, evaluations, agent workflows, grounded outputs, and clean failure paths.
Gerald Cuder
Building practical AI products with calm, focused engineering.
I build practical AI products, machine learning features, and backend systems with a focus on clarity, product fit, and reliable implementation.
About
Thoughtful systems for AI, ML, and software that still need to feel usable.
I care about shipping things that are technically serious, product-aware, and legible for the next iteration.
ML features that fit the product
The job is not just to make a model work in isolation. It is to make the entire user flow clear, trustworthy, and maintainable once real usage starts.
Engineering with product judgment
I like software that reads well, scales sensibly, and leaves room for iteration instead of locking a team into unnecessary complexity.
Milestones
Where I've done the work.
RO
Rocket Apps / StepsApp
ActiveSenior Machine Learning Engineer
Building ML features and LLM-powered experiences for fitness and health apps used by millions. Work spans backend services, iOS integration, statistical modeling, and LLMOps infrastructure.
LLMOpsiOS DevelopmentFitness AIHealth StatisticsBackend APIsData Science
SY
Syndena
Senior Machine Learning Engineer
Applied NLP and federated learning to Electronic Health Records under strict privacy constraints. Designed human-in-the-loop pipelines and LLMOps workflows that kept clinicians in control of model outputs.
NLPFederated LearningElectronic Health RecordsPrivacy-preserving AILLMOpsHuman-in-the-Loop ML
IT
iTranslate
Senior Machine Learning Engineer
Developed and shipped neural machine translation and speech recognition models to production across iOS and Android. Handled the full pipeline from research to on-device deployment.
Neural Machine TranslationSpeech RecognitionCoreMLTFLiteiOSAndroidTensorFlow
Selected Projects
A small set of work shaped around usefulness, clarity, and durable implementation.
Selected work across evaluation systems, health-oriented product concepts, and knowledge infrastructure.
Open Source Library
SoftAdaptX
A Keras-native wrapper for adaptive loss weighting based on task difficulty. Implements and extends the SoftAdapt algorithm, enabling dynamic rebalancing of multiple loss terms in multi-task and multi-output learning. Supports 2D and 3D outputs — including token-level NER — and is compatible with TensorFlow 2.x training loops.
PythonKerasTensorFlowNumPyPyPI
What I Build
The kinds of systems I like to work on.
Presented simply: a few focused areas where product thinking and implementation quality need to move together.
AI/LLM Prototypes
Fast, credible experiments that test usefulness early and still respect production constraints.
ML Features for Apps
Classification, retrieval, ranking, and decision support integrated into real interfaces and workflows.
Backend/API Development
Clear contracts, background jobs, data models, and service layers that hold up under iteration.
Evaluation and Optimization
Benchmarks, prompt iteration, latency tradeoffs, and instrumentation that turn guesswork into feedback loops.
Frontend Systems
React and Next.js interfaces that stay close to the product — clean data flows, sensible state, and UIs that hold up as features grow.
iOS Apps
Native Swift and SwiftUI development with CoreML model integration, on-device inference, and the full path from Xcode to App Store.
Get in touch
Open to thoughtful conversations about AI products, ML features, and product engineering.
gerald.cuder@icloud.com