Shivanath Devinarayanan · Updated
On this page
Route the model to the job. A time-series foundation model is not a chatbot. A long context window is not a knowledge graph. An open edge model is not a reason to skip the gateway. The map is public SDKs and model families, wired through the control plane so spend and policy stay in one place.
Google TimesFM is a decoder-only series model for zero-shot forecast across granularities. Gemini-class models take long documents, codebases, and multimodal recordings in one window. Gemma and CodeGemma stay inside a perimeter when code cannot leave. Claude Agent SDK and OpenAI SDK cover tool loops, structured outputs, and schema checks before a write. LangGraph and LangChain hold cyclic state, supervisor-worker graphs, and human-in-the-loop interrupts. Orchestration choices are on guided determinism .
Why it matters
A single frontier model for every step is how token bills become architecture. Summarization, classification, and extraction can use a fast cheap model. Legal, financial, and architectural judgment can use a frontier model. The gateway arbitrates. I run this across 200+ production agents.
Schema reflection is a loop: the model emits JSON, another pass checks it against the schema, and only then does the durable plane commit. HITL interrupts pause LangGraph until Slack records a human allow.
In a generic org
Token burn on the gateway is a series. TimesFM forecasts the month before a loop trips the commit. A 200-page MSA goes to Gemini-class long context, not a 500-token chunker. Internal code completion stays on Gemma in the VPC. A deal-risk brief uses Claude or OpenAI tool loops against a live CRM snapshot. None of those calls skip Flex Gateway.
SDK jobs
- TimesFM. Forecast run-rate and capacity from gateway telemetry and pipeline velocity. Not a generative chat surface.
- Gemini-class long context. Contract sets, audit logs, and repository-scale questions without pretending a vector store is the archive.
- Gemma / CodeGemma. Perimeter inference for code and other data that must not leave.
- Claude Agent SDK and OpenAI SDK. Multi-turn tools, structured outputs, routing to cheaper models for high-volume text.
- LangGraph. Cyclic graphs, consensus among workers, and a hard pause before a mutation.
Stack
- TimesFM
- Gemini
- Gemma
- CodeGemma
- Claude Agent SDK
- OpenAI SDK
- LangGraph
- LangChain
How to hire
Hiring starts with a governed production agent on one documented job, then a consultation.
Bring one documented job. Start at stand up production AI agents , then book a consultation . Agentforce is the Salesforce runtime: Agentforce that stays in production .
Home · About · Services · Portfolio · Success Stories · Writing · News · Events · Contact · Media kit · Privacy Policy