Meridian

Grid-flexibility software for data centers.

This repository no longer publishes the live site. As of 2026-09-18, meridian.kianfoshee.com is served by the Vercel project ercot-flex from ~/ercot-flex/site, so the site and its /api/contact route share one origin. GitHub Pages and the docs/ export are retired: editing this tree will not change what visitors see. Make site changes in ~/ercot-flex/site and deploy with vercel deploy --prod.

Meridian is a decision-support system for data-center demand flexibility. It forecasts when grid conditions may tighten, evaluates a facility's approved operating constraints, and produces a response plan for an operator to review.

System architecture

flowchart LR
  A[Grid, market, and weather data] --> B[Calibrated risk forecasting]
  B --> C[Versioned output contract]
  D[Facility telemetry and constraints] --> E[Facility flexibility model]
  C --> F[Decision engine]
  E --> F
  F --> G[Human-approved operating plan]
  G --> H[Measurement, verification, and audit trail]

Loading

What I built

• An ERCOT forecasting pipeline for physical grid tightness from one to four days ahead.
• A facility model covering load, cooling, storage, workload, contract, and recovery constraints.
• A human-in-the-loop console for planning, event response, and post-event verification.
• ERCOT and LADWP market adapters behind one versioned output contract.
• Point-in-time validation, explicit abstention, immutable releases, and hash-chained evidence ledgers.

Stack

Layer	Technology	Why it was chosen
Product interface	Next.js 16, React 19, TypeScript, Tailwind CSS	A typed UI with server and client rendering in one codebase
Hosting	Vercel for product deployments; GitHub Pages for this showcase	Preview deployments, edge delivery, and reproducible static publishing
API	Python, FastAPI, Pydantic	Typed contracts and a small, testable service boundary
Data pipeline	Polars, PyArrow, Parquet, DuckDB	Efficient analytical processing of large, time-indexed archives without operating a database server
Modeling	scikit-learn; PyTorch for research experiments	Interpretable baselines, calibrated gradient boosting, and controlled comparison with sequence models
Data sources	ERCOT public data, NOAA GFS/GEFS, Open-Meteo, EIA, and public utility records	Primary-source grid, weather, and market inputs with explicit publication vintages
Operational state	Append-only, hash-chained JSONL ledgers	Reproducible histories of inputs, forecasts, decisions, and corrections
Development tools	Claude Code and OpenAI Codex	Implementation assistance, test generation, and adversarial review

Engineering decisions

• No random train/test split. Grid events cluster in time, so models are evaluated with walk-forward folds and data as it existed at each forecast cut.
• No silent imputation. Missing or stale inputs produce an explicit degraded or abstaining state.
• No LLM in the decision path. Operational probabilities come from calibrated statistical models. AI coding tools were used for implementation and review.
• No automatic control. The console is advisory; an authenticated operator remains responsible for approval.

Current status

Historical validation is strongest for broad physical grid tightness. Operator actions are much rarer, so those estimates remain provisional. Facility-specific instructions and deliverable megawatts cannot be validated without a design partner's records. Meridian is therefore operated in shadow mode while prospective evidence accumulates.

Public repository scope

This repository contains the public product interface and a technical overview. The forecasting pipeline, model artifacts, facility optimizer, customer connectors, operational configuration, and partner data are private.

Run the public interface

pnpm install
pnpm dev

Then open http://localhost:3000.

Before committing, run the complete quality gate:

pnpm check

This runs linting, TypeScript validation, formatting checks, and a production build.

---

Built by Kian Foshee.