Services
The work, in five sections.
One page, because the practice is one practice. Each section says what is hard, what we deliver, how we approach it, and the technologies we will actually use.
01
Agentic AI and LLM applications
The problem
A useful assistant is a stateful system. It has to retrieve from your data, show its work, and refuse an answer it cannot support.
What we deliver
Agent workflows, retrieval services, guardrail and explainability layers, and an evaluation of those behaviors.
Approach
We turn the product into explicit states and tools, test the paths, and keep a trace a reviewer can read. AI coding assistants sit inside that loop. They do not replace the tests.
Technologies
- Python
- FastAPI
- LangGraph
- LangChain
- PostgreSQL
02
Production ML platforms and MLOps
The problem
A model that only the author can retrain is not a platform. The next change has to be repeatable, and a bad change has to be stoppable.
What we deliver
Containerized services, workflow platforms, and release pipelines with regression checks.
Approach
Package the job, put evaluation on the path to production, and keep that path for the next change.
Technologies
- Docker
- Kubernetes
- Kubeflow
- SageMaker
- Lambda
- S3
- CodePipeline
03
Cloud-native backend and event-driven systems
The problem
Other teams have to run the service. That means a contract, a data model that matches the queries, and a stream when the work is a stream.
What we deliver
Service APIs, event-driven workflows, and the database work those systems depend on. On a high-traffic platform we have also migrated a core service.
Approach
Small services, explicit contracts, asynchronous pipelines, and tests at the unit and integration boundary.
Technologies
- Python
- FastAPI
- Kafka
- PostgreSQL
- Docker
- Kubernetes
- AWS EKS
04
Computer vision and scientific imaging
The problem
Volumetric electron microscopy needs segmentation a scientist can run, retrain, and export, not a one-off notebook.
What we deliver
Plugin APIs, 2D and 3D panoptic segmentation, training and fine-tuning, tiling, label cleanup, and export, with tests and a release path.
Approach
We treat the scientific workflow as software: an API, a test suite, documentation, and continuous delivery.
Technologies
- Python
- PyTorch
- napari
05
Reliability, observability, and test engineering
The problem
Production AI fails in quiet ways: a regression, a slow request, a system whose health nobody can see.
What we deliver
Test-driven development, automated evaluation gates, regression suites, and dashboards for latency and failures.
Approach
Tests and evaluation are part of the change. Observability is in place before the incident, aimed at the questions an on-call engineer asks.
Technologies
- pytest
- CI/CD
- Prometheus
- Grafana
- Loki
Start with the constraint that matters.
Tell us the system, the deadline, and who has to live with the result.
Start a conversation