We built Exosphere (exosphere.host) – a platform to orchestrate and run batch AI agents on large data with connectors, autoscaling, and affordable inference (up to 75% cheaper). Think of it as a control plane for async AI workloads.

Why we built it: Running background AI workflows (like summarising 1M support chats or processing 10K PDFs) is messy – you need queueing, scaling, model hosting, cost control, and integration with your systems. Most infra today is optimised for chat apps, not bulk tasks or pipelines. This is going to become messier with multistep AI agents/workflows coming in.

What Exosphere does: - Supports batch AI agents with parallelism, retries, and memory - Integrates with tools like S3, Notion, GCS, Pinecone etc - Works with open-source models like DeepSeek, LLaMA, and Claude via API - Has a soon-to-be open-source orchestrator called Orbit (built from scratch) - Cost-optimised infra tuned for large data inference and delayed inference

Easy onboarding, no GPU setup required

Example use cases: - Classify or extract info from 100K PDFs - Run retrieval-based QA across millions of records - Summarize and route large volumes of tickets or feedback - Batch label images or text for finetuning

We’d love feedback from this community – thoughts on dev experience, connectors to add, model support, or features you'd want in the agent platform.

You can try it out here or just reply here if you want some free credits for trying open-source models in batch.

Thanks! – Nivedit (ex-Azure OpenAI) and the team