Agentic systems, evals, voice, and medical AI, engineered to production.
Every engagement follows the same three moves.
We scope the highest-leverage problem, define what “better” means, and build the evals before we build the system. No slideware: week one produces measurements, not documents.
A working system reaches production in weeks, wrapped in the engineering it needs to stay there: observability, CI/CD, durable execution, cost and latency control.
Feedback loops, eval-gated agents, and automated retraining keep the system climbing after we hand it over. The goal is a system that gets better without us, and a team that knows exactly how it works.
One accountable senior lead per engagement, drawing on a network of specialist engineers when the problem calls for it. Everyone who touches your system has shipped AI to production before.
Four areas, one standard: production, not prototype.
Multi-agent architectures, tool use, and the harness engineering that makes agents dependable: durable execution, human approval gates, bounded autonomy, agent-legible codebases.
Pydantic AI · Anthropic/OpenAI/Gemini APIs · MCP · Temporal · FastAPI
Evaluation frameworks covering every component, used for model selection, regression prevention, and hill-climbing performance. Plus the unglamorous halves of quality: cost optimization and latency budgets.
Custom eval harnesses · Braintrust · Promptfoo · LangFuse · W&B
Realtime speech-to-speech agents on telephony and messaging, and the model layer beneath: TTS/STT training pipelines, speaker recognition, inference acceleration.
LiveKit · OpenAI Realtime · NVIDIA NeMo · PyTorch
Imaging models for clinical research and the pipelines around them, built for environments where data is sensitive and mistakes are not an option.
MONAI · PyTorch · DICOM/NIfTI · privacy-aware architecture
Underneath all four: cloud-native platform engineering. Kubernetes, GitOps, infrastructure as code, and MLOps on GCP/AWS/Azure.
Systems we've designed and built. Described without names; happy to go deeper in conversation.
A multi-channel AI assistant with realtime voice, in production, where agents triage, plan, implement, review, and ship every feature, with one human gate between a spoken request and deployed code. The same loops run onboarding, changelogs, and growth.
An internal platform of reusable, provider-agnostic components (agents, RAG, speech-to-speech, observability, CI/CD) built for a US AI consultancy; one later deployment shipped through configuration alone.
Automated training, evaluation, and deployment for the conversational-AI core of a scale-up, running unattended in production; retraining stopped needing an engineer at all.
A 3D imaging model that finds bone in scans at a fraction of the usual radiation dose, unlocking new downstream analyses in nuclear-medicine research; a second model automated fat quantification for clinical studies.
“We're extremely pleased with Gradient Ascent's support and guidance in setting up our GitOps, CI/CD and Kubernetes workflows. We've been fortunate to have them on our team.”
Our engineers have built alongside frontier AI labs, shipped for enterprises in media, energy, and healthcare, and run production systems from startup scale to 10,000+ tenants.