Adam BlansettSenior Full-Stack & AI Engineer
Featured WorkActive Open-Source Development (Apache 2.0)

Autonomous AI Virtual Assistant

Privacy-first, self-hostable autonomous virtual assistant with local LLMs and mobile approval gateways.

Engineering RoleLead Systems & AI Engineer

01. Problem & Context

Executives and founders lose hours daily managing calendar conflicts, triaging emails, taking meeting notes, and compiling briefing dossiers. Cloud AI assistants lock users into recurring subscriptions while transmitting sensitive executive schedules and emails to third-party servers.

02. Engineering Solution

Built a self-hostable autonomous assistant capable of executing multi-step workflows (scheduling, inbox zero, travel itineraries, briefing dossiers) using local open-weight models (Qwen 2.5, Llama 3.3) and Telegram mobile approval gates.

03. System Architecture

The assistant pairs a Next.js 16+ reactive executive dashboard with dynamic agent execution planners. Tasks are decomposed into inspectable JSON tool execution graphs, verified against security policy gates, and executed against live Google Calendar and email APIs with post-execution state verification.

Core Architectural Components

Local Open-Weight Model Adapter

Connects to local Ollama servers running Qwen 2.5, Llama 3.3, and Mistral models with structured JSON schema enforcement and zero cloud telemetry.

Dynamic Multi-Step Agent Planner

Decomposes natural language requests into structured tool chains, piping outputs into downstream inputs with policy guardrails.

Telegram Mobile Gateway with Inline Approvals

Ingests mobile requests and voice memos on Telegram, routing consequential mutations to interactive inline approval buttons ([✓ Approve] / [✕ Reject]).

Persistent Long-Term Memory

Securely remembers executive preferences, VIP contacts, travel habits, and recurring routines, injecting context into planning prompts.

Google Calendar API v3 Coordinator

Manages meetings, enforces 15-minute buffer periods between calls, resolves scheduling conflicts, and compiles pre-meeting intelligence dossiers.

04. Notable Engineering Decisions

Local-First LLMs via Ollama with Frontier Fallback

Executive correspondence and calendar data should never leave the local machine by default. Local models (Qwen 2.5 14B) handle daily routines, falling back to frontier APIs only when explicitly enabled.

Mobile Telegram Approval Gate for Consequential Actions

Autonomous systems cannot be allowed to send emails or cancel meetings without confirmation. Interactive Telegram buttons ensure 100% human-in-the-loop oversight.

Comprehensive Test Harness (131 Passing Tests)

Agent systems with multi-step tool calls require exhaustive regression testing across plan generation, schema validation, and credential isolation.

05. Engineering Challenges & Solutions

Challenge: Preventing SSRF attacks during automated web research and URL ingestion.

Engineering Solution: Engineered a dedicated network sandbox blocking loopback addresses, internal subnets, and cloud metadata endpoints before fetching untrusted web pages.

Challenge: Handling token context limits when processing messy meeting transcripts.

Engineering Solution: Implemented map-reduce document summarization extracting action items, deadlines, and attendee sentiment without exceeding model context windows.

06. Implementation Highlights

  • Local open-weight model integration with Ollama (Qwen 2.5, Llama 3.3, Mistral)
  • Multi-step dynamic agent chaining with JSON schema validation
  • Telegram mobile gateway with instant inline approval buttons
  • Direct Google Calendar API v3 synchronization with automatic buffer guarding
  • AES-256-GCM authenticated encryption for OAuth tokens and API credentials
  • 131 automated tests passing across planning, execution, and security layers