Not concept decks. Agent loops, client portals, verification pipelines, and VPS infrastructure — with guardrails, tests, and honest boundaries.
Live private beta
Estate Mogul OS
Telegram ops agent with workspace actions, not chatbot replies.
A production assistant loop for operators: Telegram intake, intent routing, LLM reasoning, Google Workspace tools, preview-confirm before outbound actions, and VPS deployment with regression gates.
Problem
Operators needed an assistant that could read deal flow, draft email copy, and route workspace tasks — without blind automation or unsafe connector claims.
Approach
Built a layered runtime: deterministic routing and guardrails first, Hermes brain for open-ended reasoning, separate OAuth stores per connector, and human confirmation before Gmail send or sheet mutation.
What shipped
Telegram polling worker with attachment and caption handling
Google Workspace CLI with scoped OAuth and token isolation
Preview-confirm queues for sheet updates and Gmail send
Connector honesty lane when scopes or tokens are missing
systemd service on Linux VPS with validate/smoke test gates
Service-user isolation for secrets and runtime state
Python · Telegram · Hermes · Google Workspace · Gmail · Sheets · systemd · Linux VPS · pytest
Live client app
Knox Client Portal
Self-hosted client portal with autosave, admin APIs, and CRM sync.
Token-based onboarding portal for a luxury real-estate client program — questionnaire flows, branded email briefs, Zapier sync, and migration from managed DB to self-hosted SQLite on VPS.
Problem
A live client portal needed reliability, ownership of data, and operational control without depending on a paused managed database host.
Approach
Rebuilt the persistence layer on SQLite, added health checks and backup cron, deployed behind nginx with systemd, and kept Zapier as the integration edge.
What shipped
Autosaving multi-step questionnaire with submit flow
Admin APIs and client record management
Zapier sync cron and branded Resend email briefs
nginx + systemd production deploy on VPS
Runtime health endpoint and verification scripts
Guardrails
Token-scoped client access
Server-side validation on all questionnaire writes
Queue-based lead verification with enrichment and operator visibility.
Backend pipeline for validating and enriching lead records — Redis-backed jobs, third-party enrichment, proxy-aware processing, and dashboard visibility for operators.
Problem
Lead lists needed verification and enrichment at scale without manual spreadsheet work or opaque batch failures.
Approach
Designed a queue-first architecture with BullMQ, isolated integration adapters, and observable job states for debugging and recovery.
What shipped
Job queue with Redis and worker processes
Email validation and enrichment integrations
Operator dashboard for pipeline visibility
Dockerized deployment path
Guardrails
Rate limits and retry boundaries on external APIs
Job failure surfaces with actionable states
Secrets kept out of client bundles
Next.js API · BullMQ · Redis · Docker · Apify · Apollo
Private infrastructure
AI Model Runtime
Dockerized model routing, workflow automation, and eval gates.
Self-hosted stack for routing LLM traffic, running workflow automation, and holding production data locally — LiteLLM, n8n, Postgres, Redis, and validation runbooks.
Problem
Needed a controllable AI runtime on owned infrastructure instead of scattering prompts across ad hoc SaaS tools.
Approach
Composed core data services, LiteLLM for model routing, n8n for workflow scaffolding, and shell-based validation gates before any workflow activation.
What shipped
Docker Compose core + workflow profiles
LiteLLM bridge with local-only bindings
n8n workflow imports with disabled-by-default posture
