Technology

Paramus need a advanced level of digital transformation

If you need assistance with an assessment, we can evaluate your current maturity level (consulting) and recommend next steps to move forward in your digital journey.

Architecture

The current version has a Router / Multi Agent layout. PARAMUS uses LangGraph. Most agents are ReAct type.

Run on different LLM providers

The CORE as well as the agents need their own LLM runtimes subscriptions! This costs are additional to PARAMUS but give you the freedom to balance your spendings. You have the choice how much power PARAMUS will have in which field of expertise.

Mix them up for a well balanced system

You should configure different runtimes for different agents! Just like your team members have different skills (and salaries). This is just as comparable: By assigning agents to providers differently, we have achieved incredible results in optimizing the performance of PARAMUS (as a whole system) in terms of response time, response quality and total costs.

For example: run the free Calculator agent with non-critical, cheap but VERY FAST(!) xAI-Grok2 and run in contrast the PARAMUS Core with a openAI-GPT4o. To our experience the PARAMUS CORE (supervisor) should run with the best model you can effort. Simple agents like the calculation Agent can run „more stupid“.

Runtime LLMs available:

• OpenAI
• Google Vertex AI (untested)
• XAI
• Anthropic

There are already LLMs with agent infrastructures. So why PARAMUS?

(1) Paramus ‚thinks‘ Chemistry, the agents and tools have prompts that are designed with chemistry in mind. General architectures know a lot – but they need to be tuned for our specific field of expertise and workflow. It is a challenge to design prompts for relevant use cases in chemical industry and test them all over the agent ecosystem. Its no so much about the technical framework, Its about „the orchestration of chemistry in it“.

(2) Agent performance excellence and answer quality due to Knowledge Graph technology, see Knowledge Graph

(3) Simple and easy setup, you are not bound to a big infrastructure vendor or have to go into the cloud.

(4) Basic projects and task management that is helpful for scientific work. „Just lists of chats“ is not enough for productive work with GPT’s.

(5) PARAMUS keeps you up-to-date with the latest agentic technology. There is a dynamic change in this area: speed matters. A cohesive ecosystem can be deployed more quickly. There are two releases per year: March and October.

SECURE by design

How to install PARAMUS

Safety philosophy

Every user has his own PARAMUS. No shared environments – this is by design! The reason is, that PARAMUS is your personal assistant, really powerful and so need to know the access to the systems on your behalf. Shared environments may be safe meanwhile … but still not enough for us to trust.

Do you offer agents without the whole infrastructure?

Each agent is LangGraph-compatible, so basically we can do it. But you will loose the power of the whole PARAMUS graph, the End-to-End Value vs. Single-Use Component. If you integrate an agent you need to figure out the “who calls what, when, and how?”. You also have to manage the meta-logic of chaining multiple agents, handle edge cases, and state management. In our current version we use a ‚orchestrator-worker‘ architecture for PARAMUS Core.

In a system like PARAMUS, each agent might be specialized for a certain domain (e.g., experiments/ELN, time series/reactors, synthesis route handling and call for prediction tools etc.). Because they’re embedded in a larger framework, each agent can pass information to other agents via the supervisor. This synergy means the whole is greater than the sum of the parts.

When multiple agents work under a shared framework, logs, user feedback, and performance metrics can be aggregated across tasks. This consolidated feedback can improve each agent over time and help the orchestrator (PARAMUS) learn (agent) routing strategies or identify bottlenecks.

We recommend to use PARAMUS Itself Can Act as an Agent:

PARAMUS as meta-agent

From the outside, a customer or client application just sees a single interface (the “PARAMUS agent”). Internally, PARAMUS handle the sub-agents. You do not need to wire all these sub-agents yourself; with python simply pip install paramus and call paramus.runTask() endpoint, for example, and let the PARAMUS chemist figure out everything else.