EN RU
SYSTEMS ARCHIVE

Selected Systems

Architectural case studies and R&D projects built around hard constraints: memory budgets, deterministic evaluation, air-gapped security, and high-throughput inference.

SYSTEM / 001 AI MEMORY R&D ACTIVE
2026

Skinki — Local-First Memory Engine

A portable, local-first memory engine designed for small on-device models (~4B parameters on Apple Silicon M1 Air). Explores how hierarchical vector compression, sleep-time consolidation, and provenance graphs make personal AI agents truly context-aware without remote API leakage.

RUST MCP RETRIEVAL EVALS EMBEDDINGS MEMORY C-ABI/FFI
5M vectors <250 MB RAM (RaBitQ 1-bit / Int4)
0.291 → 0.438 Recall@10 (Coarse-to-fine hierarchy)
0.325 → 0.800 Synthetic Graph Win (Pass gate)
0.168 vs 0.193 Real Dialogue Falsification Signal
[ CASE STUDY & ARCHITECTURE ] RESEARCH NOTE → GITHUB ↗
SYSTEM / 002 REALTIME / ON-PREMISE HARDWARE APPLIANCE
2026

Vnutri — Hardware Appliance Messenger

An on-premise B2B communications system shipped as a dedicated physical ARM appliance (Rockchip RK3528A / RK3588S). Features zero-knowledge E2EE cryptography (X25519 DH + AES-256-GCM) shared across React PWA (WASM) and Swift iOS (UniFFI), write-buffered SQLCipher storage (260x fsync drop), and an on-demand AI sidecar socket.

RUST AXUM WEBSOCKETS ARM SQLCIPHER WASM UNIFFI REACT SWIFT
~130 MB App RAM (500 active concurrent users)
260x Drop Fsyncs via memory write-buffering
33-Byte Binary Header (O(1) blind routing)
Zero-Knowledge Dual-payload sender/recipient sync
[ CASE STUDY & ARCHITECTURE ] HARDWARE TIERS → DECISION MATRIX →
SYSTEM / 003 AI OPERATIONS & AGENTS PRODUCTION / ACTIVE
2026

Philipp — Ambient AI Project Manager

An ambient AI project manager that listens to team chatter without conversational spam, maintains a structured Kanban board, and dispatches validated tasks to an autonomous coding execution loop (Antigravity CLI) with 0 API tokens spent on routine hygiene.

PYTHON GEMINI 3.8 FLASH SQLITE / WAL ANTIGRAVITY CLI TELEGRAM BOT API WHISPER V3 TURBO SYSTEMD TIMERS
0 API Tokens Deterministic Systemd Hygiene & Deadlines
Issue → PR End-to-End Autonomous Coding Loop
15–20 / wk Sustained Weekly Triage Throughput
12s Debounce Single-Pass Dual Decision (PM vs Speech)
[ CASE STUDY & ARCHITECTURE ] DISCUSS DEPLOYMENT →
SYSTEM / 004 TALENT INTELLIGENCE & EVALUATION PRODUCTION / ACTIVE
2026

HR Screener — Automated Interview Intelligence

An asynchronous candidate screening and interview intelligence engine that scrubs Russian PII and conversational noise locally before cloud transit, yielding ~35% token compression and grounded competency evaluations in <8s at ~$0.001 per screening.

GO (ECHO V4) ASYNQ / REDIS POSTGRESQL DEEPSEEK-CHAT HMAC-SHA256 PII SANITIZER
PII Masked Before Cloud Local Pre-LLM Regex & Lexical Anonymizer
~35% Saved Token Compression via Noise Stripping
< 8s Turnaround Asynchronous Go/Asynq Worker Pipeline
~$0.001 Inference Unit Cost (DeepSeek V4 Flash)
[ CASE STUDY & ARCHITECTURE ] DISCUSS DEPLOYMENT →
SYSTEM / 005 CONTENT INTELLIGENCE & SYNDICATION PRODUCTION / ACTIVE
2025–2026

Orion — Autonomous Content Syndication Engine

An unattended multi-channel content intelligence and publishing engine that uses local dense vector embeddings (e5-small) to filter out cross-outlet near-duplicates, extracts high-res hero media, and publishes structured analytical briefs under systemd supervision.

PYTHON 3.12 SENTENCE-TRANSFORMERS E5-SMALL NUMPY GPT-4O-MINI SYSTEMD TIMERS
0.92 Gate Local Dense Vector Cosine Filter (e5-small)
< 900 Chars Strict Editorial Contract ("Why It Matters")
Runs Unattended Systemd Timers with Active-Hour Windows
W × H Scored Deep-Page Hero Media Area Resolution
[ CASE STUDY & ARCHITECTURE ] DISCUSS DEPLOYMENT →