What Buffaly is
Buffaly is an open-source, neurosymbolic engine designed to make AI safe, controlled, and explainable. Instead of trapping your workflows in unpredictable "text loops" like standard LLM agents, Buffaly strictly separates AI reasoning from code execution. You interact in natural language, but the actual work is executed safely through typed objects and native code.
Who this page is for
Read this page first if you are a new user trying to understand the core philosophy behind Buffaly before you start building. It answers “what is Buffaly?” and “why shouldn't I just use a normal chatbot API?”
Buffaly is built for developers and operators who need AI to perform high-stakes administrative work (like healthcare revenue cycles, back-office automation, or data staging) where hallucination is unacceptable and auditability is a strict requirement.
The short version: The 5 Pillars
To understand Buffaly, you just need to understand these five core concepts that replace the standard AI playbook:
What it actually does in practice
Because Buffaly connects AI to concrete systems, typical tasks are highly structured and evidence-driven.
- Automate insurance compliance, eligibility checks, and billing safely.
- Clean and resolve inbound patient data without exposing it to the LLM.
- Call your existing internal APIs securely through typed runtime tools.
- Normalize incoming files (like EOBs) and route exceptions automatically.
- Generate 100% durable audit trails for every single automated decision.
A real-world example
Instead of asking a standard agent to "read this 10k token JSON of patient data and process the claim" (expensive, insecure, prone to injection), you ask Buffaly to "Process this batch." Buffaly maps this intent to your native ClaimProcessor.RunNativeBatch(), executing it locally at zero token cost and zero risk of data leakage.
Buffaly vs. Standard Agents
The AI industry is treating large language models as control planes—reducing reality to giant prompts and hoping for the best. Buffaly structurally separates the reasoning from the doing.
| Standard Chat Agents (The Text Loop) | Buffaly (Native Bound) |
|---|---|
| Rely on ever-growing context prompts to know what to do. | Promotes repeated learning into permanent executable code tools. |
| Sensitive info (like SSNs or API keys) must be pasted into the context window. | Raw data stays in local memory; the AI only sees safe Reference Handles. |
| Hallucination is always a risk because execution is text-based guessing. | Actions map to strictly typed native functions (e.g., C# classes). |
| Vulnerable to Prompt Injection (user inputs overriding instructions). | Injection is structurally blocked because business logic is native code. |
| Costs scale linearly or exponentially as you do more work. | Token costs drop by ~80% as software takes over repeated tasks. |
The Core Mental Model
When using Buffaly, think of the LLM purely as a translator of intent, while your runtime is the executor.
- Intent: You ask for an outcome in plain language.
- Translation: Buffaly uses language reasoning to identify the correct typed action, tool, or entity for the job.
- Execution: Buffaly triggers your native, pre-written business logic to do the work.
- Observation: The system logs the outcome with perfect auditability.
- Optimization: If this pattern repeats, Buffaly creates a structured shortcut so the LLM doesn't even need to figure it out next time.
How it learns: Memory and Extension
Buffaly doesn't just "remember" by appending text to a chat transcript. It uses ProtoScript—a code-like language that structures memory logically. As Buffaly experiences repeated workflows, it shifts knowledge across these categories:
Why safety and verification matter
In heavily regulated fields like healthcare and finance, you cannot hand the keys of your infrastructure to a probabilistic model. Buffaly is built to ensure you never have to.
- Deterministic Logic: The AI chooses the path, but the path itself is strict, reviewed code.
- Audit Trails: Every action is mapped, logged, and ready for regulatory review out of the box.
- No Improvisation: The chatbot is never left to "guess" how to handle edge-case operational logic.
- Privacy by Default: Operational evidence is handled without leaking secrets.