What counts as custom ML here?
Not a Kaggle model on a laptop. Custom ML is the ranking, retrieval, speech, or lip-sync path that has to meet a latency and permission budget inside an existing product. Integrations are the unglamorous half: OAuth2 to Workday, JWT rooms for LiveKit, Stripe webhooks, SOAP where REST is not enough.
Proof from systems that already exist
The ecommerce recommender exposes ranking into a storefront workflow over a private catalog index. ChainTech’s advisory chat sits on chaintech.co with FastAPI retrieval behind a Next.js workspace. The Voice Avatar Platform mints LiveKit JWTs from a Flash handler, runs MuseTalk on GPU, and never ships the model to the browser. Peoplix speaks Workday SOAP and REST with token refresh.
How we integrate ML into a live system
- Inventory the edges. List the APIs, auth, data contracts, and SLAs the new model path must respect. Guessing a greenfield rewrite is how projects stall.
- Budget latency and cost. Decide what must be GPU-local, what can be an API, and what is cached. Avatar lip-sync and live meeting Q&A have different budgets than batch ranking.
- Contract first. Ship an API or event contract your existing app can call—ranking, retrieval, room tokens—before polishing the model notebook.
- Observe in their tools. Health, CUDA or queue depth, and failed writes should show up where operators already look, not only in our staging dashboard.
What you receive
- An integration contract (API, webhook, or room token) into the existing product
- The model or retrieval path that meets the agreed latency and cost budget
- Auth that matches the host system (OAuth2, JWT, HMAC)
- Operational hooks: health, logs, and a rollback story
- Source and runbooks so your engineers can extend the edge
Related work and reading
Frequently asked questions
Do you train foundation models from scratch?
No. We compose retrieval, ranking, speech, and GPU inference on top of models and vendors that already exist, then integrate them into your product. Training a new foundation model is a different business.
Can you integrate with Workday or other HRIS SOAP APIs?
Yes. Peoplix uses Workday SOAP and REST with OAuth2 token refresh, identity lookup, and audited writes. Older enterprise APIs are normal; we do not require a greenfield GraphQL estate.
What about GPU workloads in the browser?
We do not ship MuseTalk to the client. The Voice Avatar Platform keeps inference on GPU (for example RunPod), streams frames through LiveKit, and uses a serverless handler for tokens and CUDA health.
Will you wrap a vendor API and call it a product?
A vendor is a component. The product is the contract, the grounding, the operator UI, and the write path. If the work is only a key in an env file, we will say so in discovery.