About
An engineering studio for AI systems that go to production.
We are a small team of senior engineers building ML platforms for e-commerce and medicine. We architect, we ship, and we operate what we deliver.
What we do
Four lines of work.
- Architecting AI systems for e-commerce and medicine
- Designing intelligent pipelines that power product recommendations, demand forecasting, and clinical decision-support tools.
- Multi-model LLM orchestration
- Building orchestration over Claude, GPT, and Gemini behind a unified Kosong abstraction, with Pydantic-validated schemas across the stack.
- Custom models for image, video, and voice
- PyTorch-based models for modalities that off-the-shelf APIs cannot cover — calibrated, evaluated, and operated in production.
- Engineering leadership
- Leading engineering teams in delivering production-grade ML platforms at scale — from architecture decisions through to on-call rotations.
Approach
Four phases. No theatre.
- 01
Architect
Map the problem domain, model the data, decide where ML adds value and where it does not.
- 02
Validate
Prototype, benchmark, measure. No model in production without a baseline beaten on real data.
- 03
Ship
Pipelines with Pydantic-validated schemas, observability, and rollback. CI/CD as a habit, not an exception.
- 04
Operate
Monitor drift, retrain on a cadence, keep cost per inference under control. We stay when the data shifts.
Principles
Four things we believe.
- Evidence over enthusiasm
- We will tell you when ML is the wrong answer.
- Interfaces over vendors
- Models are commodities. The interface that orchestrates them is not.
- Boring infrastructure
- Postgres, Pydantic, predictable deploys. The interesting parts live upstream.
- Operate what we ship
- If we deliver a model, we expect to wear the pager for it.
Start