Agentic AI lab building autonomous systems for fintech and real-time B2B sales operations.
We design and ship multi-agent AI infrastructure where it matters most: production trading systems that compound capital, and live sales workflows that compound revenue. Original work, shipped end-to-end.
Real-time AI coaching platform for B2B insurance sales. 12 specialized agents deliver sub-800ms coaching cues during live calls — compliance alerts, objection handling, plan matching, and tonality feedback. Self-healing portal integration via three-layer extraction (CDP network interception → accessibility tree → vision-language model fallback).
- Landing page: portal-three-rose.vercel.app
- Chrome extension: Chrome Web Store →
- Status: Public beta, validating with early-access agencies
- Source: proprietary — commercial product
Full coaching sidepanel — Sales Coach v1.6.0 docked next to SunFire's Medicare enrollment portal. Nine specialized cards (Enrollment, Script, Notes, Plan Match, Research, Ancillary, Rebuttals, Tonality, Compliance) update in real time during the call.
 |
 |
| Automation Lab — agent-driven actions on a live carrier portal. Interview, Pick, Send, Approve, Deny, Capture, Save as Workflow. | Live script with phase tracking. Opener → Problem → Presentation → Agitation → Quality. Verbatim opener rendered for the agent to read. |
 |
 |
| Interview Mode — agent-led multi-choice prompts replace freeform instructions. The agent asks structured questions to configure each new portal workflow. | Carrier portal integration — the "Sales Coach" started debugging this browser banner confirms live CDP-level network interception on the Medicare enrollment portal. |
Autonomous quantitative trading infrastructure. Full Rust + Python stack with multi-agent strategy evolution, alternative-data pipelines (satellite imagery, AIS shipping, regime / sentiment classification), and Nomad-orchestrated deployment. 30+ MCP servers and 15+ specialized agents.
- Source: proprietary — commercial product
Autonomous AI web agency platform. A coordinated agent swarm takes a client brief and delivers a deployed, monitored website end-to-end. Intake agent extracts brand guidelines, browser operator generates designs via headless Playwright, copywriter QA enforces tone, debug engineer validates in sandbox, and MCP deployer ships to Shopify or Wix through a Rust gateway. Four-tiered memory (Postgres, Redis, Mem0/Qdrant, Neo4j) ensures agents learn from every engagement.
- Source: proprietary — commercial product
The reusable building blocks underneath the products above. All original work, all production-tested.
| Repo | What it is |
|---|---|
| site-mapper-agents | LLM-driven self-healing API discovery — Pydantic AI agents (Architect, Eavesdropper, Healer) adapt to schema changes |
| arch-viewer | MCP-native codebase analysis — interactive architecture diagrams, Neo4j knowledge graph, 17 Claude Code tools |
| cdp-network-interceptor | Chrome DevTools Protocol network capture for MV3 extensions — PII redaction, iframe auto-attach, stale-debugger recovery |
| mv3-audio-replay-buffer | Encrypted, durable audio frame buffer for Chrome MV3 service workers — ack-based replay over WebSocket |
| devkit | Universal B2B/SaaS development foundation — skills, hooks, agents, CI/CD templates |
Languages — Python · Rust · TypeScript · Shell
AI / Agents — Multi-agent orchestration · MCP servers · LLM routing (Ollama Cloud, OpenAI, Anthropic) · RAG (Qdrant, Neo4j)
Backend — FastAPI · async WebSockets · Supabase Postgres · Redis · ClickHouse · Stripe
Browser — Chrome MV3 · CDP network interception · accessibility-tree crawling · VLM fallback
Frontend — Next.js 15 · Tauri 2 · responsive sidepanel UIs
Infra — Nomad · Railway · Cloudflare · Docker · CI/CD automation
Norman Beckford — Licensed Medicare Agent who taught himself to build the software his industry wouldn't. Started learning to code in 2018, building on LLMs since 2022.