Privacy-first AI infrastructure
Architectures shaped for on-premise, VPC, hybrid, and edge deployment, where data residency and auditability cannot be compromised.
Founder & Systems Architect
Donald Filimon
Building privacy-first AI infrastructure — durable memory, traceable retrieval, and bounded autonomy — close to the metal across Zig, Rust, and Swift.
United States
Builds privacy-first AI infrastructure, developer tools, and high-performance systems across Zig, Rust, Swift, TypeScript, and GPU-oriented runtimes. Leads WDBX retrieval, the Abbey–Aviva–Abi orchestration framework, and the ABI runtime.
Focus areas
Vector & memory systems
WDBX — durable vector/block stores for traceable retrieval and long-lived agent memory, implemented across several languages.
Polyglot systems engineering
Working close to the metal across Zig, Rust, Swift, and TypeScript: databases, GPU-oriented runtimes, CLIs, and developer tooling.
Compiler & low-level optimization
AI-driven compiler optimization and MLIR/Zig systems work — tuning the layers beneath the model, not just the model itself.
Agent orchestration & safety
The ABI runtime and the Abbey–Aviva–Abi persona system for bounded, observable autonomy with explicit operator control.
Signature work
abi
An AI agent runtime paired with the WDBX substrate — high-performance local orchestration in nightly Rust, with GPU capability reporting and an MCP server. The Zig tree has been removed.
Rust
WDBX
A durable vector/block memory store for traceable retrieval and agent memory, with implementations spanning Zig, Rust, Python, and TypeScript.
Rust · Python
gama
A Swift + MLX framework for on-device language-model inference on Apple Silicon, including a self-learning agent.
Swift
Nyon
A voxel 3D world experiment exploring a novel hexa-gravity system and creator-first game development.
Zig
Building close to the metal
Donald is a polyglot systems engineer who works deliberately low in the stack. His public projects span Zig, Rust, Swift, TypeScript, and Python — databases, GPU-oriented runtimes, agent frameworks, and the tooling that holds them together — because the guarantees Quesar cares about (latency, provenance, data residency) are won or lost at that level.
That range is intentional, not scattered. Each language is chosen for where it pays off: nightly Rust for the current ABI runtime and sibling WDBX substrate, Swift and MLX for on-device inference on Apple Silicon, and TypeScript for the operator-facing layers. Earlier Zig implementations remain historical context, not current ABI build guidance. The same instinct runs down to the compiler itself — projects like Cellstrap extend Zig toward MLIR, in line with his stated focus on AI-driven compiler optimization.
Memory you can trace
The throughline across his work is WDBX — a durable vector and block memory store designed so retrieval is traceable rather than opaque. The ABI runtime plans queries and assembles context packs on top of it, and the same store backs the long-lived memory that agents reason over.
The design commitment is simple to state and hard to honor: no retrieval claim without a traceable source or confidence signal. That principle is what makes the infrastructure suitable for private, audited deployments instead of demos.
Autonomy with a boundary
On top of retrieval, Quesar's orchestration layer adds the Abbey–Aviva–Abi persona system — Abbey the empathetic polymath, Aviva the direct expert, and Abi the adaptive router — operating under one motto: "Care first. Clarity always. Competence throughout."
It is built around bounded execution and explicit approvals: no autonomous write action without an observable policy boundary, and escalation, review, and override flows kept visible so the people responsible stay in control. Safety before scale is the founding constraint, not a later addition.
Building something aligned?
Start an inquiry