Back to Product Home Systems Engineering for AI

Rust-Native Runtime.
Deterministic Autonomy.

Most AI agent frameworks are built with bloated Python libraries, leaking memory and generating unpredictable execution loops. Steffi takes a systems approach: a compiled Rust core managing local quantized routing engines, memory layers, and headless browser sessions natively.

Core Technology Pillars

Discover the low-latency, memory-safe primitives that allow Steffi agents to run 24/7 background tasks with high predictability.

ONNX runtime selection gateway

Custom-Trained Local Brain (SLM)

Agents use a lightweight, specialized Small Language Model (SLM) running locally on the worker node. This model handles memory compaction and tool routing, selecting a maximum of 5 tools and 1 skill per message turn.

By shifting routing decisions to a local model, we eliminate high-latency remote calls and reduce API costs by up to 5x compared to standard agent platforms.

Headless Rust browser integration

Specter Native Browser Engine

Built directly in Rust, the Specter browser enables agents to navigate single-page applications (SPAs) and execute web workflows natively and safely. Runs isolated without external browser processes.

Note: Specter skips WebGPU support to maintain lightweight execution, preventing complex WebGPU-based fingerprint or mining scripts from running.

OTP-style Process Supervision

Supervised DAG Scheduler

Task loops run as Directed Acyclic Graphs (DAGs) under strict process trees modeled after Erlang-OTP. If an agent thread gets stuck, the supervisor restarts it from its last memory hash.

Timezone-aware scheduling handles intervals, events, and cron rules with built-in missed-run recovery.

Episodic, Semantic, Short, & Long Context

4-Tier Persistent Memory

Instead of simple history logs, Steffi utilizes a hybrid context engine. It saves episodic event paths, semantic domain guidelines, short-term context, and long-term key-value fact cards.

Vector database embeddings index the memory store, automatically loading the correct historical context in milliseconds without prompt bloating.

Performance Benchmark

Engineered to Minimize Token Overhead

Standard Python-based frameworks pass entire chat histories and massive tool lists directly to external LLM APIs on every turn. This leads to prompt bloating, tool confusion, and soaring operational costs.

5x Fewer TokensLocal ONNX model selects the necessary tools, feeding only relevant skills to public LLM tokens.
Sub-12ms Browser StartupsHeadless Rust Specter engine spins up sandboxed browser sessions in milliseconds, not seconds.

Architecture Comparison

Feature Parametersteffi.ai (Rust Engine)Python Frameworks (LangChain/CrewAI)
Base LanguageCompiled Rust (optimized binary)Interpreted Python
Token EfficiencyUp to 5x fewer active tokensBaseline (large full-history context)
Background ExecutionHours or days (fault-tolerant DAG)Seconds or minutes (memory leak prone)
Memory FootprintSub-45MB base worker memory250MB - 1GB+ runtime dependency memory
Security GuardrailsBuilt-in 6-layer scanner binaryBasic regex filters (easily bypassed)
Package SandboxingNative Unix Socket gRPC docker gatewayLocal shell execution (highly risky)
Modular Workspace

The Rust Crate Structure

rs

rb-core

Engine orchestrator, memory writers, and task execution registry.

rs

rb-agent

Actor-model execution context, quantized LLM router, and council deliberation modes.

rs

rb-plugin

SaaS connections, capability permissions, and MCP client implementations.

rs

rb-guardrail

Prompt injection scanners, data loss prevention (DLP) masking, and text normalization.

rs

rb-api

Dual-auth REST API endpoints and real-time WebSocket subscription channels.

rs

rb-scheduler

Fault-tolerant cron scheduler and interval trigger handlers.

rs

rb-storage

redb and PostgreSQL hybrid engine with active CDC synchronization.

rs

rb-sdk-gateway

Per-run Unix Domain Socket (UDS) gRPC gateway for sandbox script environments.

Fun Fact

The codebase internal prefix "rb" stands for "rusted-batata". In honor of sweet potatoes, the team developed batata-clean memory cycles to purge and compress vector caches dynamically!

Our Leadership

Built by Olib AI

Steffi is engineered and maintained by Olib AI. Our team combines deep systems engineering, enterprise compliance, and AI research backgrounds.

Tyonna M.

CEO

Former Vice President of Operations at JP Morgan Chase. Tyonna directs corporate compliance, financial operations, and growth strategy.

Akram H. S.

Founder & CTO

Previously Lead Software Developer at Aloha. Akram oversees core development, Chromium engine compiler patches, and low-level modules.

Connect on LinkedIn →

Maya M.

Chief Architect of AI

Maya directs on-device model deployment, llama.cpp fine-tuning, element matcher architectures, and semantic task processing.

Olib AI (founded 2024) is headquartered in Stone Mountain, Georgia, United States. Contact: support@olib.ai | sales@olib.ai