Advection Software: Real-Time LLM Streaming & Edge Architecture
[!NOTE]
Executive Overview
Attribute Details Role & Tenure Founding Product Engineer (2023 – Present · 27+ months) Company Focus AI-native product lab, multi-agent developer tooling, real-time media streaming, autonomous product synthesis. Core Mission Architect resilient multi-agent orchestration workflows, headless verification harnesses, context engineering networks, and user-facing generative AI engines. Tech Stack Node.js, TypeScript, Cloudflare Workers & D1, Serverless FFmpeg, Google GenAI SDK, Playwright, Zod Key Highlights Multi-agent orchestration over custom remote gateways, headless verification harnesses, node-based context engineering networks, user-facing dynamic media/software/metadata engines.
At Advection Software, Andrea operates as the Founding Product Engineer, leading full-stack architectural design, multi-agent systems orchestration, execution harnesses, and edge infrastructure.
1. Multi-Channel LLM Streaming Protocol
Bridging generative agent loops with responsive user interfaces requires deterministic communication contracts:
- Structured Event Envelopes: Replaced raw text streams with strongly typed JSON event envelopes (validated via Zod), streaming tool calls, UI component triggers, thinking deltas, and telemetry concurrently.
- Backpressure & Client Hydration: Designed reconnection protocols that replay unacknowledged event streams upon temporary network drops, ensuring state synchronization between client and serverless runtimes.
- Low-Latency Streaming: Optimized server-sent events (SSE) and WebSocket transports to stream tokens with sub-40ms time-to-first-byte (TTFB).
2. In-House Edge Ad Network (Cloudflare Workers + D1)
To cross-promote internal products without third-party ad network tracking bloat or cookie banners, Andrea architected a bespoke edge house-ad system:
- Zero-Cold-Start Global Delivery: Runs on Cloudflare Workers edge nodes across 300+ locations, executing impression tracking, placement auctions, and creative selection in under 5ms globally.
- Lightweight D1 Storage: Utilizes SQLite-based Cloudflare D1 databases at the edge with periodic batch reconciliation to primary data stores.
- Strict Format Constraints: Enforces compact WebP animated previews, AV1 video clips, and WebM assets with strict pre-fetching headers, avoiding mobile layout shifts.
3. Serverless Video Transcoding & Ingestion
- On-Demand FFmpeg Pipelines: Automated serverless video processing using Google Cloud Functions and AWS Lambda containers equipped with FFmpeg.
- Adaptive Bitrate Transcoding: Converts recorded high-framerate gameplay captures into H.264, WebP, and OGV renditions with normalized audio tracks.
- Automated Social Clipping: Programmatically crops and burns animated captions for automated cross-platform promotional posting.
4. Agentic Workspace Automation & Cost Governance
- Multi-Step Agent Execution Loops: Built autonomous 15-step agent loops capable of reading multi-file projects, applying targeted code diffs, and verifying file integrity.
- Micro-Dollar Usage Tracking: Implemented tiered token accounting that tracks prompt, completion, and native thinking tokens in real time, enforcing per-user budget guardrails.
5. Multi-Agent Orchestration & Headless Execution Harnesses
- Remote Gateway Orchestration: Directing multiple concurrent AI agents across custom remote gateways (
supervise.js,term-web) over secure overlay networks, eliminating local environment bloat and isolating workloads. - Headless Verification Harnesses: Wrapping AI agents in headless execution harnesses (Playwright, Linux virtual framebuffers, AST linters) to autonomously run, test, and verify code diffs before merging.
- Node-Based Context Networks: Authoring portable, decentralized knowledge graphs that provide structured context and architectural constraints directly into agent prompts without token bloat.
6. User-Facing Generative AI Engines
- Autonomous Multimodal Synthesis: Engineered production AI engines driving end-to-end synthesis of dynamic media, playable software, and rich multi-tier metadata (e.g., AudioTales script/voice dramatization, GameTsunami playable p5.js game synthesis, Fluenics pronunciation assessment).
- Metadata Generation & Guardrails: Structured outputs producing rich scenario metadata, phoneme scoring matrices, and state envelopes that bridge AI inference directly with reactive client interfaces.