← Visual Explainers
Portfolio

What Got Built

AI platforms, evaluation engines, governed legal systems, voice pipelines, workflow automation, and 132 visual explainers — built across six domains, deployed to production.

6
Production Platforms
50+
AI Agents Built
1,050+
Tests Passing
132
Visual Explainers
10+
Deployments
100+
API Endpoints
Platform 1
AESOP Transformation OS
Full-lifecycle AI agent development platform. Discover → Design → Build → Evaluate → Repair → Track. The evaluation engine that later powered Legal-OS's Harvey monitoring layer.
50 agents · 75 API routes · 76 migrations · ~990 tests
The Evaluation Engine
The hardest problem in AI: how do you evaluate an AI agent when the agent itself cannot be trusted to self-assess? AESOP answers with 20+ specialized evaluation agents, each scoring a different dimension — then a programmatic scoring engine applies weighted tiers, veto rules, and certification levels. No LLM self-assessment. Every score is auditable.
20+ evaluation agents 6-stage pipeline Scoring: 4 dimensions + veto Cave of Shadows (red-team) The Hydra (parallelized swarm) Bias Probe (differential testing) Sentinel (QA test gen) Drift Detection (6 types) RAG Help System (pgvector) Custom Rubric engine Black Box Investigation Forge Repair agent
6-stage pipeline from discovery through production monitoring
Scoring engine: Tier 1 (1.5×) + Tier 2 (1.0×), hard veto at 75
Cave of Shadows: adversarial red-team with attack personas
Hydra: parallelized swarm running dozens of test scenarios simultaneously
Bias Probe: paired differential testing detecting treatment disparity
Sentinel: generates dual-output QA test plans from evaluation results
Drift Detection: 6 drift types, weekly sweeps, severity alerts
Build pipeline: 6-phase automated agent generation with HITL checkpoint
Portfolio layer: department Kanban, strategic prioritization agent, CSV imports
Monday.com integration: MCP client, direct PRD endpoint, argument mapping
Backend: FastAPI + Supabase + Celery | Frontend: Next.js 16 + TypeScript
Deployed: Fly.io (backend) + Vercel (frontend) | ~990 tests passing
Platform 2
Legal AI Operating System
A governed platform for building, deploying, measuring, and independently evaluating AI across the legal enterprise. Built specifically for the K&I Program Manager role at Ashurst Perkins Coie.
9 functions · 10 API routes · 9 migrations · 62 tests · 11 help docs
Governance Architecture + Harvey Monitoring
Five-layer operating model: Knowledge Foundation → Governance & Trust → AI Functions → Program Operations → Organizational Model. The breakthrough: independent Harvey agent evaluation — ported from the AESOP scoring engine, adapted for legal outputs. Harvey can't grade its own homework. Legal-OS does.
9 AI functions deployed Harvey Agent Monitoring 4-dim evaluation (ported from AESOP) ROI Framework: time × rate POC Pipeline (Kanban) Agentic Help System (RAG) Portfolio Dashboard Enablement Kit (7 docs) RLS client isolation ABA 512 compliance ready
5-layer architecture: Knowledge, Governance, Functions, Operations, Org Model
Harvey evaluation: external agents pulled in, 4 parallel scorers, weighted + veto
ROI Framework: cost avoided = time saved × billable rate from rate cards
POC Pipeline: Kanban board — Discovery → Build → Review → Graduated/Cancelled
Agentic help system: Voyage embeddings → pgvector → RAG answers (89 chunks)
Audit trail: append-only, immutable, every LLM call recorded
Provider resilience: tries Anthropic first, auto-falls-back to DeepSeek
Backend: FastAPI + Supabase + Celery | Frontend: Next.js 16 + TypeScript
Deployed: Fly.io (backend) + Vercel (frontend, legal.sickofancy.ai)
52 backend tests + 18 E2E Playwright tests — all passing
Platform 3 & 4
SuperAssistant & TSE — Voice + Workflow AI
AI assistant platform for The Bradbury Group. Voice interview pipeline, email automation, and workflow orchestration — serverless architecture across Railway, Supabase, and ElevenLabs.
SuperAssistant MVP
  • 39 API route files across backend
  • Multi-tenant architecture on Railway + Supabase
  • Agent orchestration with structured output pipelines
  • Knowledge base integration with vector search
  • Production deployment with CI/CD pipeline
TSE — Voice Interview Pipeline
  • Serverless voice interviews via ElevenLabs
  • Structured question flow with branching logic
  • Real-time transcription + analysis
  • Deployed on Railway + Supabase
  • Fully automated end-to-end interview pipeline
TBG Automations
Operational AI for a consulting firm — turning manual workflows into automated pipelines.
L&D Cohort Automation: end-to-end training program automation through Sprint 9
Gmail Automation: AI email triage using Gemini + Gmail API, research phase complete
System Instructions: central source of truth for all TBG AI system instructions
Ecosystem Map: operational overview of all Railway/Vercel/Supabase/Cloudflare deployments
Walter: fully TBG-owned AI assistant, deployed and operational
Matter Intake Evaluator: standalone deployment via Vercel from repo root
Visual Communication
132 Visual Explainers
Self-contained HTML pages — architecture diagrams, maturity models, process maps, audit reports, platform showcases — spanning every project. No frameworks. No build step. Just hand-built HTML + CSS that renders instantly.
30 AESOP diagrams 19 Legal OS diagrams TBG architecture & audits Monday.com pipeline (6 stages) Executive one-pagers Maturity models Storage architecture Resume evidence matrix Process flowcharts Platform showcases
Deployed via GitHub Pages at motorthings.github.io/diagrams
Custom visual-explainer skill: 4 templates, 3 rendering approaches, responsive design
Mermaid integration: flowcharts, sequence, ER, state machines, mind maps
Dark/light theme support on every page with persisted preference
Executive, technical, and operational views for every platform
Section navigation with sticky TOC and IntersectionObserver active states
AI Skills & Agents
Custom Claude Infrastructure
Skills, agents, and tools that extend Claude Code into a specialized development and analysis environment — each with defined prompts, knowledge bases, and evaluation criteria.
Visual Explainer Skill
Generates self-contained HTML diagrams. 4 template types, 7 aesthetic palettes, Mermaid + CSS + Chart.js. The skill that built all 132 explainers.
PRD Skill
Socratic discovery across 4 depth levels, 7 phases, 6 reference files. Generates complete PRDs from structured interviews.
Transcript Analysis
Transforms meeting transcripts into structured discovery reports. Explicit + implicit signal detection. Stakeholder mapping.
Brand Voice
Edgy, brutal-honest, poetic style. Tension/release rhythm. Applied to LinkedIn content, professional writing, and platform copy.
Thinking Partner Agent
Socratic reasoning partner. Challenges assumptions. Maps trade-offs. Structured decision analysis.
15+ Specialized Agents
Research assistant, interviewer, daily summarizer, chat archiver, project triage, career guidance, and more.
Cross-Cutting
Patterns That Span Every Project
These patterns appear across every platform — they're not one-off implementations but reusable architectural primitives ported between projects.
Multi-Agent Evaluation
  • Built in AESOP, ported to Legal-OS
  • Parallel evaluator agents — each scores a different dimension
  • Programmatic scoring engine — never trusts LLM self-assessment
  • Weighted tiers + hard veto + certification levels
  • Drift detection with baseline comparison and severity alerts
  • Same scoring formula used in both platforms
RAG Help Systems
  • Built in both AESOP and Legal-OS
  • Voyage AI embeddings → pgvector similarity search
  • LLM-generated answers with source references
  • Slide-out chat panel with suggested questions
  • Local FAQ fallback when API unreachable
  • Same architecture, same indexing pipeline
Governance Architecture
  • Append-only audit trails — every decision recorded
  • Row-Level Security — client data isolation at the database
  • Human-in-the-loop gating with configurable thresholds
  • Compliance readiness: ABA 512, SOC 2, ISO 42001
  • Traceability: full lineage from input to output
Provider Resilience
  • LLM provider abstraction — one interface, multiple backends
  • Automatic fallback on auth errors
  • Per-function provider overrides for cost optimization
  • DeepSeek for cost (~10x cheaper), Anthropic for quality
  • Pricing tables with per-model token cost calculation
Infrastructure
Technology Stack
Backend
FastAPIUvicornPython 3.12+PydanticCeleryasyncpg
Frontend
Next.js 16React 19TypeScriptTailwind CSSPlaywright
Database
SupabasePostgreSQLpgvectorRLSRedis
AI / LLM
Claude (Anthropic)DeepSeek V3GeminiVoyage AIElevenLabs
Deploy
Fly.ioVercelRailwayCloudflareGitHub Pages
Dev Tools
Claude CodeGitHub ActionsObsidianMonday.comClickUp
Six platforms. 50+ AI agents. 1,050+ tests. 132 visual explainers.
Every platform follows the same principles: governance-first, deterministic over LLM, the audit trail is the product.
Built with Claude Code by Charlie Fuller.