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 — Rust + Python, from compute kernels to browser extensions.

Real-time AI coaching platform for B2B insurance sales. 13 specialized agents deliver live coaching cues — compliance & SOA disclosure tracking, objection handling, plan matching, research, and tonality feedback — under a sub-800ms end-to-end latency budget. Self-healing carrier-portal integration: a three-layer perception stack (CDP network interception → DOM/selector mapping → vision-language-model fallback) plus an event-driven healer that auto-repairs when carrier APIs change. Real-time voice runs through a Rust (axum + tokio) WebSocket service streaming to Deepgram, with PII/PHI redaction in-Rust.

• Live: salescoach.axumquant.com
• Chrome extension: Chrome Web Store →
• Status: Public beta (v1.9.7), validating with early-access agencies
• Source: proprietary — commercial product

The coaching panel in action — Sales Coach docked beside SunFire's Medicare enrollment portal. The live script tracks call phases (Intro → Qualify → Needs → Research → Presentation → Enrollment), while call notes and permission-to-record prompts update in real time.

Plan Match + AI Research — ranked carrier plans with one-click SOB, and an agent that surfaces supplemental coverage (Vision, Dental, Hearing) mid-call.	Ancillary + Rebuttals — cross-sell prompts and real-time objection handling, from a credibility stack to no-pressure-close scripts.
Compliance + Tonality + Coaching — recording-notice & disclosure tracking, live tonality metrics, and coaching cues delivered as the call unfolds.	Modular card system — each specialist agent renders as a card the rep toggles on demand: Enrollment, Script, Notes, Plan Match, Research, Ancillary, Coaching, Compliance, Rebuttals, Tonality, Psychology, Transcript.
AI roleplay training — reps practice against AI personas and get scored on discovery, objection handling, compliance, and tonality before they ever touch a live call.	Built-in diagnostics — Deepgram STT health checks and multi-carrier portal readiness (SunFire, EnrollHere, Heartland, UHOne…).

Autonomous quantitative trading infrastructure. Rust + Python: Rust compute kernels (PyO3/maturin), island-model evolutionary strategy discovery, and a 9-stage statistical validation gauntlet (walk-forward, Deflated Sharpe, CPCV, Monte-Carlo) that rejects overfit strategies before any capital is risked. 30+ strategy archetypes, 20+ specialized agents, 30+ MCP servers, and 80+ Nomad jobs; alternative-data pipelines (satellite imagery, AIS shipping, sentiment / regime classification).

• Architecture overview: github.com/axumquant/axum-trading-platform
• 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 an MCP deployer ships to Shopify or Wix through a Rust gateway. Four-tiered memory (Postgres, Redis, Mem0/Qdrant, Neo4j) lets agents learn from every engagement.

• Architecture overview: github.com/axumquant/axum-labs-studio-overview
• Source: proprietary — commercial product

Deep architecture write-ups for the proprietary products above — no source, full design.

The reusable building blocks extracted from the products above — all original work, production-tested, each with real browsable source (not just a README).

Languages — Python · Rust · TypeScript · Shell

AI / Agents — Pydantic AI · LangGraph · multi-agent orchestration · Model Context Protocol (30+ servers, FastMCP) · LLM routing (Ollama Cloud · OpenAI · Anthropic · Groq) · RAG & memory (Qdrant · Neo4j · Mem0)

ML / Data — PyTorch · scikit-learn · XGBoost · LightGBM · Optuna · gplearn · Polars · NumPy / SciPy · CuPy (GPU)

Backend — FastAPI · Pydantic V2 · async WebSockets · gRPC · Alembic · Supabase / Postgres · Redis · ClickHouse · Stripe

Rust — axum · tokio · PyO3 / maturin (native Python extensions) · sqlx · fred · tokio-tungstenite · thirtyfour

Browser — Chrome MV3 · CDP network interception · DOM/selector mapping · Playwright · VLM fallback

Voice — Deepgram (STT) · ElevenLabs (TTS) · Rust WebSocket pipeline · in-Rust PII redaction

Frontend — Next.js 16 · React 19 · Tauri 2 (Leptos / WASM) · Prisma · shadcn/ui

Observability — Arize Phoenix (LLM tracing) · OpenTelemetry · Sentry · structlog

Infra / DevOps — Nomad · Vault · Cloudflare · Docker · Railway · Vercel · GitHub Actions

Norman Beckford — Licensed Medicare Agent who taught himself to build the software his industry wouldn't. Coding since 2018, building on LLMs since 2022.