Tools & Instruments
Live dashboards, simulation engines, and AI-powered instruments. Each tool is a working artifact — not a mockup.
Real-time AI inference monitoring — latency, cost, throughput.
A full-featured monitoring console for AI inference workloads. Tracks token streams, compression ratios, router distribution, and per-call cost savings across 4 active routes. Displays live tok/s meter, 12-minute stream chart, and a ledger of recent inference calls.
Financial view of accumulated inference savings via router API.
Connects to the SubtracToken router API to display all-time accumulated inference, token reduction rates, actual vs. baseline cost, and a full financial breakdown including naive, same-model, billable, and customer-net savings. Requires router URL and API key.
Bi-weekly competitive intelligence + self-correcting benchmark loop.
Runs every other Monday at 12:00 UTC. Benchmarks TokenEase against LiteLLM BM25 and bear-1 (The Token Company) across 5 prompt categories. Claude Haiku judges each compression on semantic fidelity, instruction integrity, and concept survival. Detects drift vs. historical baseline and self-corrects the benchmark suite when quality drops.
Daily DeFi Analysis Report quality engine — judge, drift, self-correct.
Runs daily at 06:00 UTC. Fetches DeFi market data via Perplexity (DefiLlama, CoinGecko, macro, sentiment), generates a Bloomberg-quality report via Claude Sonnet, then judges it on signal density, data freshness, narrative clarity, and risk coverage. If verdict is REVISE, regenerates once with feedback injected. Detects quality drift and updates system prompt + source weights automatically.
Pre-execution decision engine — 10-check gate before any AI action runs.
A FastAPI router that enforces NSZ-001 through NSZ-010 checks before any agentic action executes. Evaluates environment conditions, risk inputs, and idempotency state. Maintains a hash-chain audit log via SQLite WAL. Uses optimistic concurrency and compare-and-swap to prevent duplicate execution. Risk score above threshold blocks execution with a full audit trail.
Qiskit + OpenMM + Three.js quantum chemistry simulator.
An interactive H₂ molecular dynamics simulator combining classical MD (OpenMM) with quantum energy evaluation (Qiskit VQE). Drag hydrogen atoms to change bond length, run molecular dynamics, evaluate the quantum Hamiltonian expectation value, and scan the θ landscape. Renders in Three.js with live energy trace chart. Falls back to demo state when heavy engines are unavailable.
Closed-loop infinite optimizer — rulial engine with ABO mutation tracking.
A canvas-based cellular automaton running a modified Rule 30 with ABO (Adaptive Behavioral Optimizer) mutation loop. Tracks intelligence score, joy baseline, and mutation parameters (evolveFactor, hueDrift, breathAmp, ruleBias) in localStorage. Mutates every 45 seconds, accepts or prunes based on simulated joy gain. Six visual feature modes including Aura (Feature 5) with heart rate modulation.
AI receptionist vs. agent cost modeling across call volume and workflow count.
A cost comparison model for AI receptionist vs. AI agent deployments. Models incremental cost across 50–1,000 calls/month and 0–10 workflow integrations. Calculates break-even call counts and the additional calls needed to justify agent value at each workflow tier. Based on real pricing data from Marcie AI deployments.
Several are available for licensing or integration. Others are open for collaboration.
Get in Touch