Tim Tuev
01 / Executive Summary
I turn business and product problems into functional AI systems — from model selection and inference hardware strategy to vector substrates, agents, long-term memory, retrieval, and deployment. Over nearly a decade in product design and creative leadership combined with deep hands-on AI systems R&D, I connect business viability with technical execution, pairing with coding agents for rapid implementation while owning architecture, experiments, evaluation, and product trade-offs.
02 / Core Competencies
AI SYSTEMS & R&D
- Model Selection: Open weights vs proprietary APIs, Dense vs MoE architectures, context budgeting.
- Inference & Quantization: llama.cpp, MLX, vLLM, GGUF/AWQ/EXL2, KV cache optimization, MTP speculative decoding.
- Memory & Retrieval: Local-first memory substrates, mmap vector indices, RaBitQ 1-bit compression, hierarchical coarse-to-fine search.
- Evaluation & Evals: Empirical benchmarking (LongMemEval, LoCoMo), synthetic ground-truth generation, statistical FDR filtering.
- Agent Architectures: MCP (Model Context Protocol) tool calling, deterministic context assembly, zero-knowledge sidecars.
SOFTWARE & INFRASTRUCTURE
- Backend & Crypto: Rust (Axum, Tokio, deadpool), X25519 DH, AES-256-GCM, PBKDF2/Argon2, SQLCipher WAL.
- Cross-Platform: Rust compiled to WebAssembly (WASM) for browsers and UniFFI for native Swift/iOS.
- Hardware Appliances: Low-power ARM silicon (Rockchip RK3528A/RK3588S), eMMC wear write-buffering (~260x fsync drop).
- Frontend & UX: TypeScript, React 19, Vite, Tailwind CSS, Virtualized UI (react-virtuoso), Framer Motion.
03 / Selected Systems & Track Record
Designed and built a portable, local-first memory engine in Rust enabling small ~4B parameter models on Apple Silicon to perform multi-hop contextual reasoning without remote cloud leakage.
- Achieved 5,000,000 vector index residency in <250MB RAM via RaBitQ 1-bit binary quantization with Int4 reranking.
- Designed hierarchical coarse-to-fine search boosting recall@10 by 46% (0.291 → 0.438) over standard dense baselines.
- Conducted and published empirical research on GraphRAG transfer falsification on LongMemEval dialogue.
- Built Model Context Protocol (MCP) server bindings for seamless agent tooling integration.
Architected a turn-key physical ARM appliance (Rockchip RK3528A / RK3588S) for air-gapped corporate real-time communication.
- Developed shared cross-platform cryptographic core (`vnutri-crypto`) in Rust, compiling to WASM for Web and UniFFI for Swift/iOS.
- Designed 33-byte binary header MsgPack wire protocol enabling O(1) blind message routing by the server without payload parsing.
- Implemented in-memory write-buffering for SQLCipher, reducing disk fsyncs by ~260x to protect appliance eMMC flash storage.
- Optimized full application memory to ~130MB RAM under 500 concurrent connections, fitting within a 2GB hardware ceiling.
Founded an independent long-form analytical publication breaking down business models, strategic positioning, design systems, and tech architectures for a ~7K audience of CEOs, founders, and investors.
Nearly a decade leading multidisciplinary design teams, creating complex brand architectures, digital wayfinding schemes, enterprise product workflows, and zero-to-one product user experiences.