Federico Kamelhar

so excited to see this out in the wild! 🎉 got to team up with @LiteLLM to make OCI Generative AI a first-class provider — Llama, Grok, Cohere, Gemini & gpt-5 through one OpenAI-compatible gateway, signing handled for you 🚀 #LiteLLM #OCI #GenAI x.com/OracleCloud/st…

LiteLLM now supports Oracle Generative AI Infrastructure as a first-class provider. social.ora.cl/6012BDuEx6 You can route requests to models hosted on Oracle Generative AI Infrastructure, while LiteLLM handles OCI Signature v1 request signing and centralizes production controls.

Had the privilege of working on this 🙏 @LiteLLM now natively supports Oracle Generative AI Infrastructure — route Meta Llama, xAI Grok, Cohere, Gemini & OpenAI gpt-5 through one OpenAI-compatible gateway, OCI signing handled for you. blogs.oracle.com/ai-and-datasci… #litellm

New Haystack integration: Oracle AI Vector Search. Use Oracle Database 26ai as your vector store in any Haystack pipeline — native VECTOR type, HNSW + IVF indexing, hybrid search (vector + keyword), and metadata filtering. Ships with OracleDocumentStore, OracleEmbeddingRetriever, and OracleBM25Retriever — drop them into your indexing and query pipelines. Haystack for the RAG pipeline. Oracle for the vectors. Big thanks to David Batista (@davidsbatista on GH), @Julian_Risch, and @LukawskiKacper for the collaboration. haystack.deepset.ai/integrations/o… #Haystack #RAG #VectorSearch #OracleDatabase #Oracle26ai #GenAI #LLM #AI

@DarshanSays What that typing buys: in n=200 adversarial across 4 judges, removing the four-way partition drops proceed by up to −49; ρ=0 inflates score on every judge. GSAR decides ship/rewrite/replan; the action layer reads the partition. (2/2)

@DarshanSays @DarshanSays Appreciate the safety lens. Quick reframe: GSAR isn't gating tool calls — it evaluates *claims in the produced summary*. §1 example: "CPU >95% at t" → tool_match (w≈1); "likely caused the spike" → inference (lower w). Same sentence, different evidence type. (1/2)

I’ve just shared a new paper on arXiv: arxiv.org/abs/2604.23366 GSAR: Typed Grounding for Hallucination Detection and Recovery in Multi-Agent LLMs As more teams build multi-agent systems for enterprise AI, a practical question arises: How do you know the system is actually pointing you in the right direction? Most current approaches reduce this to a single score or pass/fail signal. However, what’s missing is a method to turn that signal into actionable insights for the system. This paper aims to make that shift by treating grounding not just as evaluation, but as a control layer for agentic systems, which includes: - Proceed - Regenerate (a cost-effective fix) - Replan (more costly, but sometimes necessary) The goal is to advance from simply asking “is this grounded?” to understanding what kind of grounding signal we have and determining the appropriate actions to take. Key insights from this work include: - Not all evidence is equal; tool outputs versus model inference are more significant than expected. - Small design choices, such as how contradictions are handled, can greatly impact outcomes. - Different models exhibit varied behaviors when tasked with judging grounding. If you’re working on or deploying agentic systems, I’d be interested in your perspective — and feel free to take a guess at which model “wins” before reviewing the results. #AI #LLM #Agents #EnterpriseAI #MachineLearning #OracleCloud #Anthropic #OpenAI

Any developers building AI applications on OCI Generative AI? We’ve got good news! social.ora.cl/6015BBMROc You can run the same Cohere code on OCI Generative AI—including Command A, Command A Vision, Command A Reasoning, Embed v4, and Rerank 3.5—while keeping authentication, compartment governance, and inference within your OCI tenancy.

Great to see this out in the wild! Had a fantastic time collaborating with Daniel Ohashi and Michael Kozakov from Cohere on this. Now, customers on Oracle Cloud Infrastructure can tap into Cohere’s power through a native SDK—unlocking faster, more seamless AI integration at scale. #GenAI #AI #CloudComputing #OCI #Cohere #DeveloperExperience #MachineLearning #EnterpriseAI

We’re incredibly excited about our partnership with OpenAI and remain focused on building and delivering the capacity they need to support rapidly growing demand. We’re seeing firsthand how quickly adoption of their technology is accelerating, driven by the strength of their latest models. OpenAI’s new 5.5 model is a significant step forward, and we expect continued momentum as access to their technology expands across cloud providers. Together, we’re enabling customers to bring powerful AI capabilities into production at scale.

Excited to share: the @cohere SDK is now natively integrated with @OracleCloud Generative AI — making it simpler for developers to build with Cohere models directly on OCI. Grateful to have collaborated with the Cohere team on this. blogs.oracle.com/ai-and-datasci… docs.cohere.com/docs/oracle-cl… #OCI #GenerativeAI #Cohere

@OracleDevs Great to see this guide published! 🚀. I had the pleassure to partner with vLLM team. OCI is an official path for Enterprise Inference deployment on OKE. @OracleCloud @OracleDevs @vllm_project

This tutorial walks you through deploying the vLLM Production Stack on OKE—from infrastructure provisioning to running your first inference request. social.ora.cl/6012hNgEp

@OracleCloud @OracleDevs Official langchain-oracle samples: github.com/oracle/langcha…

I had the privilege to collaborate with @langchain on the official docs for langchain-oracle and OCI. Glad to see it live. @OracleCloud @OracleDevs docs.langchain.com/oss/python/int… #LangChain #OCI

@MistralAI @OracleCloud #MistralAI #Mamba #GPU #Inference #LLMInfra

Excited to contribute upstream to @MistralAI’s official mistral-inference repo with improved interactive support for Mamba models. Validated on real NVIDIA GPU hardware with Codestral-Mamba. @OracleCloud github.com/mistralai/mist…

Official OCI vLLM Production Stack : github.com/vllm-project/p…

Excited to see Oracle Generative AI included in the official vLLM Production Stack on OCI. This Oracle tutorial helps make that path more accessible on OKE. @vllm_project @OracleCloud docs.oracle.com/en/learn/deplo… #vLLM #OCI #OKE