Hi, I’m Roni.
I build autonomous systems.
Lead systems & AI engineer at RS Systems. I architect autonomous agent platforms, credential-isolated proxies, APIs, and mission-critical Linux infrastructure.
Core Production Systems
Autonomous multi-agent platforms running in production with real user traffic, automated revenue pipelines, and hardened Linux infrastructure.
Felicity
Zero-Dependency Security GateThe security boundary for autonomous agents. A pure standard-library auditing engine that reads a repository the way an attacker would: high-entropy secret detection, Shannon AST vulnerability patterns, interprocedural taint analysis, MCP and agent sandbox validation, IaC and container posture, and SARIF 2.1.0 output for CI. Bound to every push on a machine through a pre-push gate that chains a structural commit gate, so nothing dangerous leaves a laptop — or a fleet of agents — unexamined.
Helpotron
Live Production · Paying StudentsAn all-in-one AI study platform and workflow OS engineered for university students. Features an autonomous 15-category verification swarm, multi-agent MCP tool triage, and SQLite WAL persistence. Supports multimodal coursework dump (audio transcription, slides, notes), Socratic tutoring agents, double-entry token accounting, and real-time task synthesis.
T2B Qualification Engine
Autonomous Revenue PipelineAn end-to-end sales qualification and lead engagement platform for commercial contractors (roofing, solar, HVAC). Automatically ingests municipal permit ETL data, executes multi-channel AI conversational loops across Twilio SMS and Telegram with zero human touch, and routes high-intent qualified jobs directly into a paywalled contractor portal with Stripe billing integration.
Security Gates, Context Compaction & Runtimes
AI-coding-orchestrator
Multi-Lane Fix ExecutorAutonomous batch fix executor. Parses multi-file audit plans into disjoint batches, drives parallel LLM lanes with automatic failover, applies byte-exact edits, verifies in-process, and escalates stalled steps via a 3-tier solver.
webchat-to-api-harness
Agent API GatewayOpenAI- and Anthropic-compatible HTTP API gateway backed by real browser webchat sessions. Enables Aider, LiteLLM, and Claude Code to drive sessions with complete file and shell tool calling.
commit-condom
Pre-Push Security GateAir-gapped Git pre-push policy gate for AI agents. Enforces atomic commits, limits change size, and blocks secret leaks behind isolated proxies.
token-condom
Write-Time Output CompressorDeterministic write-time tool output compressor and deduplicator. Collapses repetitive logs, test runs, JSON, and diffs before ingestion into agent context to protect provider prefix caches.
tool-call-compactor
Dynamic Schema CompactorCompresses bloated MCP JSON schemas into crisp 30–100 word functional definitions with clear titles so models instantly understand tool capabilities without context bloat. Dynamically indexes and retrieves full schemas only when an explicit tool invocation occurs, eliminating schema drift, token exhaustion, and tool hallucinations across 1,100+ tools.
context-rot-detector
Real-Time Rot TelemetryAttaches directly to active agent sessions to monitor context degradation across 15 behavioral signals—including instruction amnesia, attention dilution, token bucket exhaustion, and error loops. Calculates the exact mathematical compaction knee to execute context compaction at the optimal window, maximizing cognitive coherence while minimizing token costs.
memory-condom
Prefix-Cached Persistent MemoryInjects persistent, structured memory into every agent turn while strictly safeguarding provider prefix caches. Enforces an append-only memory policy that only permits in-place rewrites immediately following compaction events, paired with rigorous word budgets and terse output constraints to permanently prevent agents from yapping their context window away.
Laya-jev-mcp
Decision Gate MCPCalibrated probability engine for agent decisions: evaluates candidate actions, returns calibrated confidence scores, and enforces act/escalate thresholds to prevent coin-flip guesses from executing.
anti-spiral.js
Loop Abort PluginStreaming reasoning loop detector: intercepts repetitive model generation in real-time, aborts looping turns within hundreds of characters rather than tens of thousands, and preserves normal code and logs.
Engineering sovereign autonomy.
I (Roni Saguey) manage the entire technical side: from systemd service daemons, multi-agent MCP frameworks, and context compression to air-gapped security gates and full-stack software development across 42 repositories.
Verified Certifications
Python Certification
ClarityCheck Innovation
Technical Communication
Let’s Connect
Reach out directly for engineering collaborations, platform integrations, or architecture questions.