Autonomous Dual-Process (System 1: TypeSafe Jev + System 2: Amazon Bedrock & LangGraph) Energy Grid & Power Factory Optimization Platform with Andrej Karpathyβs Self-Recurring Continual Learning Loop
Created & Engineered by Rishabh Patil (@MrRobotop)
About HelioGrid-AI Β· Control Room UI Showcase Β· System Requirements Β· 30-Second Quick Setup & API Keys Β· Pre-Trained Models & Benchmarks Β· System Architecture
Modern electrical transmission grids and thermal power factories operate under unforgiving nonlinear physical lawsβKirchhoff's AC power flow equations (
Existing approaches suffer from a fundamental dichotomy:
- Legacy Operations-Research & SCADA Solvers optimize static snapshots but lack semantic root-cause awareness, cross-substation anomaly reasoning, and autonomous adaptation when ambient weather or load distributions shift.
- Standard Generative LLMs ("System 2" token-by-token models) are too slow (
2β8 secondslatency) and probabilistically uncalibrated for real-time control routing, and frequently hallucinate unphysical bus voltages or violate branch MVA thermal ratings when used without hard physics guardrails.
HelioGrid-AI solves this by uniting a rigorous Physics & Mathematical Optimization Core (System 0) with a Dual-Process Cognitive Architecture (System 1 TypeSafe Jev + System 2 Amazon Bedrock & LangGraph) and Andrej Karpathy's Self-Recurring Continual Learning Loop (autoresearch):
-
System 0 β Physics Digital Twin, Hybrid ML & Constrained AC-OPF Core:
- Built on 34,840 hours of real substation telemetry (
ETTh1+ETTh2), 9,568 hours of Combined Cycle Power Plant (UCI CCPP) operations, and a full Newton-Raphson AC Power Flow Digital Twin of the IEEE 14-Bus Hybrid Transmission Network (3 non-renewable CCPP gas/steam turbines, 4 utility solar/wind DERs, and 2 grid-scale BESS units). - Forecasts multi-horizon plant power output and nodal net load using a serialized Hybrid Stacking Ensemble (
Second-Order Newton-Boosted Histogram Trees+PyTorch ResidualGridNet+Out-of-Sample Calibrated Ridge Meta-Learner) wrapped in Split-Conformal Prediction Intervals (P10 / P50 / P90) and Exact-Efficiency Shapley (SHAP) sensor attributions. - Solves constrained nonlinear AC Optimal Power Flow (AC-OPF) & BESS Arbitrage alongside CCPP Inlet Air Chilling & Condenser Vacuum Optimization via
SciPy SLSQP.
- Built on 34,840 hours of real substation telemetry (
-
System 1 β Fast Calibrated Decision Engine (
TypeSafe JevΒ·typesafe/jev-1.13):- Executes sub-200ms calibrated probabilistic decisions trained via Reinforcement Learning for Calibrated Decisions (RLCD) across three structured primitives:
Noul(binary safety/contingency gates),Choice(workflow routing), andScore(calibrated$[0, 1]$ severity metrics). - Operates at three critical control points: Route-First Triage Controller (deciding whether to invoke heavy System 2 reasoning), Verify-Before-Act OPF Guardrail (blocking unsafe dispatch setpoints and detecting Jevons Paradox rebound risks), and Verify-After Anti-Hallucination Semantic Auditor.
- Executes sub-200ms calibrated probabilistic decisions trained via Reinforcement Learning for Calibrated Decisions (RLCD) across three structured primitives:
-
System 2 β Multi-Agent Analytical Mesh (
Amazon Bedrock+LangGraph+LangChain):- Orchestrates a 7-node cyclic
LangGraphStateGraphconnecting specialist tool-calling agents (plant_thermodynamics_agent,grid_forecast_agent,optimal_dispatch_agent,karpathy_autoresearch_agent, andbedrock_executive_copilot) powered by Amazon Bedrock (Claude 3.5 Sonnet/Amazon Nova Pro) with automatic region extraction and zero-downtime failover to a Citation-Locked Physics Synthesizer.
- Orchestrates a 7-node cyclic
-
Continual Learning β Andrej Karpathyβs Self-Recurring Loop (
autoresearch+ Data Engine):- Governed by the declarative specification in
program.md, HelioGrid-AI continuously ingests streaming grid and plant telemetry, proposes structural/hyperparameter mutations, backtests candidate models in shadow mode against the AC-OPF digital twin, and enforces a strict Keep / Discard Gateβreducing composite grid inefficiency$\mathcal{J}_{\text{total}}$ by42.65%as more data is observed.
- Governed by the declarative specification in
HelioGrid-AI ships with a real-time Dark-Obsidian Double-Bezel SCADA & AI Control Room (FastAPI + interactive SVG/HTML5 engine) designed with atmospheric minimalism, live animated AC power-flow current streams, interactive ambient weather stress-test sliders, and one-click Bring-Your-Own-Key (BYOK) hot-swapping.
Panel 1 & 2: IEEE 14-Bus Newton-Raphson AC Power Flow Map & System 1 (TypeSafe Jev) Decision Telemetry
-
Left Panel β Live IEEE 14-Bus Hybrid Grid Topology: Visualizes all 14 buses, 20 AC transmission corridors with animated directional power flow and thermal loading percentages (
$\rho_\ell$ ), nodal voltage magnitudes ($V_i \in [0.95, 1.05]\text{ p.u.}$ ), and optimized dispatch setpoints across CCPP gas/steam generators (Bus 1,Bus 2,Bus 6), solar/wind DERs (Bus 3,Bus 8,Bus 10,Bus 13), and utility BESS units (Bus 4,Bus 9). -
Right Panel β System 1 (
TypeSafe Jev) Fast Decision Engine: Displays real-time calibratedNoul,Choice, andScoreprobabilities across the Route-First Triage Controller, Verify-Before-Act OPF Safety Guardrail, and Verify-After Anti-Hallucination Auditor.
-
Left Panel β CCPP Power Factory Digital Twin &
SHAPWaterfall: Monitors BraytonβRankine Net Electrical Output ($P_{\text{CCPP}}\text{ MW}$ ), thermal efficiency ($\eta_{\text{th}}%$ ), heat rate ($\text{BTU/kWh}$ ),SLSQPinlet air chilling ($\Delta T_{\text{chill}}$ ) and condenser vacuum ($\Delta V_{\text{vac}}$ ) recommendations, plus Exact-Efficiency Shapley (SHAP) attributions quantifying how Ambient Temperature (AT), Exhaust Vacuum (V), Wet-Bulb Temperature (T_wb), and Relative Humidity (RH) drive plant output. -
Right Panel β 48-Hour Grid Load, Renewable Generation & Split-Conformal P10βP90 Horizon: Plots multi-horizon forecasts from
models/grid_load_forecaster_champion.ptwith calibrated Split-Conformal P10βP90 uncertainty bands (89.11%empirical holdout coverage) alongside solar/wind generation and transformer oil temperature (OT Β°C).
Panel 5 & 6: Andrej Karpathyβs Self-Recurring Autoresearch Lab & System 2 (Amazon Bedrock + LangGraph) Copilot
-
Left Panel β Karpathy Self-Recurring Loop (
autoresearch+ Data Engine): Tracks the live monotonic decline of the composite objective$\mathcal{J}_{\text{total}}$ (1.0000 β 0.5735,-42.65%) across streaming data stages (320h β 1,040h) and logs every autonomous mutation trial (KEPTvs.DISCARDED) in a persistent ledger. Operators can trigger a new self-recurring experiment live from the UI via+1 Iteration. -
Right Panel β System 2 (
Amazon Bedrock+LangGraph) Executive Copilot: Visualizes the 7-nodeLangGraphexecution pipeline (jev_system1_triageβplant_thermodynamics_agentβgrid_forecast_agentβoptimal_dispatch_agentβkarpathy_autoresearch_agentβbedrock_executive_copilotβjev_system1_auditor) and produces citation-locked control room briefings grounded in exact physical telemetry.
HelioGrid-AI is engineered to run fast on standard developer laptops, workstations, and cloud instances without requiring a dedicated GPU.
| Component | Minimum Specification | Recommended Specification | Notes |
|---|---|---|---|
| Operating System | macOS 12+, Ubuntu 20.04+ / Debian 11+, or Windows 10/11 (WSL2 / PowerShell) | macOS 14+ (Apple Silicon M1/M2/M3/M4) or Linux x86_64 | Fully POSIX & cross-platform compatible |
| Python Runtime | Python 3.10 | Python 3.11 or 3.12 | Tested with PyTorch 2.x, SciPy, FastAPI, and LangGraph |
| CPU | 2+ vCPU cores (x86_64 or ARM64) | 4β8+ CPU cores (Apple M-series / Intel Core i5+ / AMD Ryzen 5+) | Full 7-node cycle + AC-OPF executes in ~120β180 ms on CPU |
| Memory (RAM) | 2 GB free RAM | 4 GB+ free RAM | Runtime footprint with pre-trained .pt models is ~280 MB |
| GPU (Optional) | None required (CPU-first PyTorch execution) |
Apple Metal (MPS) or NVIDIA CUDA (optional) |
Full 9,568-hour model retraining takes ~4 seconds on CPU |
| Disk Space | 350 MB free space | 1 GB free space | Includes datasets (ETTh1, ETTh2, ccpp.csv) & .pt checkpoints |
| Web Browser | Chrome 110+, Safari 16+, Edge 110+, or Firefox 115+ | Latest Chrome, Brave, Arc, or Safari | Supports hardware-accelerated SVG animations & backdrop blur |
| API Keys / Network | Optional (100% functional offline out-of-the-box) | TYPESAFE_API_KEY + AWS Bedrock Key (Bearer or IAM) |
Automatic zero-downtime fallback when offline or unconfigured |
HelioGrid-AI is designed so anyone cloning the repository can launch the live Control Room and plug in their own API keys in under 30 seconds.
Run the included quickstart.sh script (or make open). It automatically creates .venv, installs dependencies, bootstraps .env, verifies the pre-trained .pt checkpoints in models/, and launches the Control Room at http://127.0.0.1:8000:
git clone https://github.com/MrRobotop/HelioGrid-AI.git
cd HelioGrid-AI
# One command: installs environment, verifies models, and opens http://127.0.0.1:8000
./quickstart.sh --open-browserOr using make:
make open # Setup + launch Control Room UI and open browser
make start # Setup + launch Control Room UI on http://127.0.0.1:8000git clone https://github.com/MrRobotop/HelioGrid-AI.git
cd HelioGrid-AI
python3 -m venv .venv
source .venv/bin/activate
pip install --upgrade pip
pip install -e ".[dev]"
# Launch the Control Room Server
python scripts/start_dashboard.py --open-browserHelioGrid-AI supports System 1 (TypeSafe Jev) and System 2 (Amazon Bedrock) credentials via three interchangeable methods:
- Start the server (
./quickstart.sh) and open http://127.0.0.1:8000. - Click the
π API Keys (BYOK)button in the top-right Floating Command Dock. - Paste your TypeSafe Jev API Key and/or Amazon Bedrock Credentials, then click
Save & Activate Keys. The status pills in the header immediately switch to live mode without restarting the server.
Copy .env.example to .env (done automatically by ./quickstart.sh) and add your keys:
# 1. System 1: TypeSafe Jev Fast Decision Engine (https://thejevai.com)
TYPESAFE_API_KEY=your_typesafe_jev_api_key_here
TYPESAFE_MODEL=typesafe/jev-1.13
# 2. System 2: Amazon Bedrock β Option I: Short-Term Bedrock Bearer API Key
# (Region is automatically extracted from the presigned token!)
AWS_BEARER_TOKEN_BEDROCK=bedrock-api-key-YmVkcm9jay5hbWF6b25hd3MuY29tLz...
# 3. System 2: Amazon Bedrock β Option II: Standard AWS IAM Credentials
AWS_ACCESS_KEY_ID=AKIA...
AWS_SECRET_ACCESS_KEY=your_secret_access_key
AWS_SESSION_TOKEN=
AWS_REGION=us-east-1
BEDROCK_MODEL_ID=anthropic.claude-3-5-sonnet-20240620-v1:0Note:
.envis strictly git-ignored via.gitignoreso your credentials are never accidentally committed.
You can inject keys directly when launching scripts/start_dashboard.py or quickstart.sh:
# Launch with an Amazon Bedrock short-term Bearer key and TypeSafe Jev key
./quickstart.sh \
--bedrock-bearer-key "bedrock-api-key-..." \
--jev-key "jev_..." \
--open-browser
# Or launch with standard AWS IAM credentials
python scripts/start_dashboard.py \
--aws-access-key "AKIA..." \
--aws-secret-key "..." \
--aws-region "us-east-1" \
--open-browserTo run a standalone end-to-end diagnostic verifying your .env credentials against live AWS Bedrock endpoints and the LangGraph workflow:
make verify-bedrock
# or: PYTHONPATH=src .venv/bin/python scripts/verify_bedrock_live.pyHelioGrid-AI ships with a hardened multi-stage Dockerfile (non-root heliogrid operator account, pre-verified .pt checkpoints, /api/health readiness probe, and per-request X-* BYOK header isolation so multiple users can safely use their own keys on a shared cloud deployment), plus one-click cloud blueprints (render.yaml and railway.json):
docker compose up --build -d
# Control Room is now live at http://127.0.0.1:8000Both hybrid forecasting ensembles (Newton-Boosted Histogram Trees + PyTorch ResidualGridNet + Out-of-Sample Calibrated Ridge Stack + Split-Conformal Quantiles) are pre-trained on a full 1-year multi-season chronicle (8,760 training hours) and evaluated on 808 unseen Year-2 holdout test hours (9,568 total hours):
| Production Artifact | Target Physical Variable | Holdout RMSE | Holdout MAE | Holdout |
Conformal P10βP90 Coverage (Target |
SHA-256 Prefix |
|---|---|---|---|---|---|---|
models/ccpp_forecaster_champion.pt (118 KB) |
CCPP Net Electrical Output ( |
0.9794 MW |
0.7927 MW |
0.9864 |
93.44% |
17fb1dc19866... |
models/grid_load_forecaster_champion.pt (141 KB) |
IEEE 14-Bus Total Active Load ( |
18.005 MW |
14.021 MW |
0.8520 |
89.11% |
d13895dbcf67... |
Full training metadata, epoch loss curves, out-of-sample stacking weights, and SHA-256 checksums are recorded in models/training_manifest.json. Re-run training anytime with make train.
Reproducible via make benchmark (scripts/run_benchmarks.py) across the IEEE 14-Bus Hybrid Network driven by real ETTh1 + ETTh2 substation active/reactive loads and UCI CCPP gas/steam turbine thermodynamics:
| Operational Metric | Baseline (Uncoordinated Droop) | HelioGrid-AI (SLSQP AC-OPF + CCPP Twin) |
Verified Improvement |
|---|---|---|---|
| Average Generation Cost | $11,415.36 / hr |
$9,876.69 / hr |
-13.48% ($258,495.41 saved / week) |
| Transmission |
13.238 MW |
9.702 MW |
-26.71% (-3.536 MW continuous) |
| Carbon Emissions Avoided | Baseline | -2,262.22 tCO2 / week |
-11.03% |
| CCPP Plant Thermal Efficiency ( |
47.76% |
48.30% |
+0.54% pts via inlet chilling & vacuum opt |
| Renewable Absorption Rate | 84.00% |
100.00% |
Zero clean energy curtailment |
| AC Voltage & Thermal Violations | 0 |
0 |
As streaming telemetry cohorts are ingested (320h β 1,040h) and autonomous hypothesis mutations run under program.md, the Keep/Discard + Jev Noul safety gate promotes genuine improvements and reverts regressions:
| Streaming Stage | Training Rows Ingested | Champion Composite Loss ( |
CCPP Output RMSE ( |
Grid Net-Load RMSE ( |
Cumulative Improvement |
|---|---|---|---|---|---|
| Stage 0 (Initial Baseline) | 320 hrs |
1.0000 |
2.050 MW |
25.12 MW |
Baseline |
Stage 1 (+180h Stream) |
500 hrs |
0.7145 |
1.465 MW |
14.98 MW |
-28.55% |
Stage 2 (+360h Stream) |
680 hrs |
0.6278 |
1.287 MW |
11.00 MW |
-37.22% |
Stage 3 (+540h Stream) |
860 hrs |
0.5902 |
1.210 MW |
9.27 MW |
-40.98% |
Stage 4 (+720h Stream) |
1,040 hrs |
0.5735 |
1.173 MW |
8.52 MW |
-42.65% (6 KEPT / 2 DISCARDED) |
flowchart TB
subgraph DATA["Layer 0: Real-World Datasets & Physics Digital Twin"]
ETT["Real Substation Telemetry (ETTh1 & ETTh2 Β· 34,840 hrs)<br/>Active Loads (HUFL, MUFL, LUFL) Β· Reactive Loads (HULL, MULL, LULL)<br/>Transformer Oil Temperature (OT Β°C)"]
CCPP["Non-Renewable Power Factory (UCI CCPP Β· 9,568 hrs)<br/>Ambient Temp (AT), Exhaust Vacuum (V), Pressure (AP), Humidity (RH)<br/>Net Output (PE MW), Heat Rate (BTU/kWh), Efficiency (Ξ·_th %)"]
IEEE["IEEE 14-Bus Hybrid Transmission Network<br/>Newton-Raphson AC Power Flow (Y_bus = G + jB)<br/>3 CCPP Units Β· 4 Solar/Wind DERs Β· 2 Utility BESS Units"]
end
subgraph SYS1["Layer 1: System 1 Fast Decision Engine (TypeSafe Jev Β· <200ms)"]
JEV_TRIAGE["Route-First Triage<br/>β’ Noul: is_emergency_contingency<br/>β’ Choice: primary_workflow_route<br/>β’ Score: grid_stress_severity"]
JEV_GUARD["Verify-Before-Act Guardrail<br/>β’ Noul: is_safe_to_dispatch<br/>β’ Noul: jevons_rebound_risk"]
JEV_AUDIT["Verify-After Semantic Auditor<br/>β’ Noul: is_physically_grounded<br/>β’ Score: operator_actionability_score"]
end
subgraph SYS0["Layer 2: Awesome-ML Forecasting, SHAP & Constrained AC-OPF"]
ENS["Pre-Trained Hybrid Forecasters (.pt Registry)<br/>Newton-Boosted Trees + PyTorch ResidualGridNet + Calibrated Ridge<br/>+ Split-Conformal P10/P50/P90 Intervals"]
SHAP["Exact-Efficiency Shapley Explainer (SHAP)<br/>+ Mahalanobis & KS Drift Detector"]
OPF["Constrained Economic Dispatch & AC-OPF (SciPy SLSQP)<br/>CCPP Inlet Chilling/Condenser Opt + BESS Arbitrage + IΒ²R Loss Min"]
end
subgraph SYS2["Layer 3: System 2 Multi-Agent Mesh (LangGraph + LangChain + Amazon Bedrock)"]
GRAPH["7-Node LangGraph StateGraph<br/>plant_thermodynamics_agent β grid_forecast_agent β optimal_dispatch_agent<br/>β karpathy_autoresearch_agent β bedrock_executive_copilot"]
end
subgraph KARPATHY["Layer 4: Karpathy Self-Recurring Loop (Autoresearch + Data Engine)"]
LOOP["Declarative Policy (program.md)<br/>Streaming Data Ingestion β Propose Hypothesis β Shadow Train & AC-OPF Backtest<br/>β J_total + Jev Noul Keep/Discard Gate β Persistent Ledger"]
end
subgraph UI["Layer 5: Dark-Obsidian Double-Bezel SCADA & AI Control Room (FastAPI + Web UI)"]
DASH["Live IEEE-14 Topology SVG Β· Jev System 1 Gauges Β· CCPP Digital Twin & Sliders<br/>48h Conformal Telemetry Β· Karpathy Autoresearch Lab Β· Live BYOK Key Modal"]
end
DATA --> JEV_TRIAGE & ENS & SHAP
JEV_TRIAGE --> GRAPH
ENS & SHAP --> OPF
GRAPH <--> OPF
OPF --> JEV_GUARD --> IEEE
GRAPH --> JEV_AUDIT --> DASH
DATA --> LOOP --> ENS & OPF
LOOP --> DASH
| Layer | Libraries & Frameworks Used | Concrete Role in HelioGrid-AI |
|---|---|---|
| System 1 Fast Decision AI | TypeSafe Jev (typesafe/jev-1.13) |
Sub-200ms calibrated Noul, Choice, and Score decisions for Route-First triage, OPF safety guardrails, and output auditing |
| System 2 Multi-Agent LLM |
Amazon Bedrock (boto3 Converse API), LangGraph, LangChain
|
7-node cyclic StateGraph coordinating specialist physics/ML agents and synthesizing citation-locked control briefings |
| Deep Learning & Boosting |
PyTorch (ResidualGridNet, .pt checkpoints), Newton-Boosted Histogram Trees
|
Multi-horizon nonlinear residual forecasting with Karpathy Data-Engine recency & tail-event sample weighting |
| Econometrics & Statistics |
Statsmodels (STL), SciPy (scipy.stats.ks_2samp, Mahalanobis distance), NumPy, Pandas
|
Seasonal-trend decomposition, finite-sample split-conformal intervals (P10/P50/P90), and covariate shift detection |
| Mathematical Optimization |
SciPy Optimize (SLSQP constrained nonlinear programming) |
CCPP inlet air chilling & condenser vacuum optimization + multi-bus Economic Dispatch & BESS arbitrage |
| Power Grid Topology & XAI |
NetworkX, Exact-Efficiency Shapley (SHAP), Newton-Raphson AC Power Flow
|
Complex nodal admittance ( |
| Control Room Backend & UI | FastAPI, Uvicorn, Pydantic, Double-Bezel SVG/HTML5 SCADA UI | Real-time REST API, interactive IEEE-14 topology map, scenario simulator, Karpathy loop lab, and Bring-Your-Own-Key (BYOK) modal |
To guarantee that multi-agent workflows never hallucinate unphysical voltages, generator outputs, or non-existent bus IDs, HelioGrid-AI grounds every agent and solver in declarative resource files:
-
program.md: Andrej Karpathy'sautoresearchdeclarative policy specifying the immutable composite objective$\mathcal{J}_{\text{total}}$ , hard physical safety invariants, and allowed hyperparameter/feature search bounds. -
resources/grid_physics_constraints.json: Exact IEEE 14-Bus voltage limits ($[0.95, 1.05]\text{ p.u.}$ ), branch MVA ratings, CCPP thermodynamic bounds, and BESS capacity/efficiency limits. -
resources/agent_prompts_and_schemas.json: Strict schema-locked prompts and System 1 (TypeSafe Jev)Noul,Choice, andScoreverification contracts. -
Deep-Dive Architecture Docs (
docs/):
HelioGrid-AI/
βββ .github/workflows/ci.yml # GitHub Actions CI (Ruff + 20 phased pytest tests)
βββ assets/ # High-resolution SVG product banners & UI panel showcases
β βββ hero_banner.svg # Product hero header & KPI showcase
β βββ panel_ieee14_topology_and_jev.svg # IEEE 14-Bus AC topology & System 1 Jev telemetry
β βββ panel_ccpp_shap_and_telemetry.svg # CCPP thermodynamic twin, SHAP & 48h conformal chart
β βββ panel_karpathy_loop_and_copilot.svg# Karpathy autoresearch lab & System 2 Bedrock copilot
βββ docs/ # Comprehensive architecture, physics & benchmark docs
βββ data/
β βββ ETTh1.csv # Real Substation 1 active/reactive load & oil temp (17,420 hrs)
β βββ ETTh2.csv # Real Substation 2 active/reactive load & oil temp (17,420 hrs)
β βββ ccpp.csv # UCI Combined Cycle Power Plant thermodynamic dataset (9,568 hrs)
β βββ benchmark_results.json # Reproducible empirical benchmark artifacts
βββ models/
β βββ ccpp_forecaster_champion.pt # Pre-trained CCPP Thermodynamic Ensemble + PyTorch checkpoint
β βββ grid_load_forecaster_champion.pt # Pre-trained IEEE 14-Bus Net Load Ensemble + PyTorch checkpoint
β βββ training_manifest.json # SHA-256 checksums, loss curves, and Year-2 holdout test metrics
βββ resources/
β βββ grid_physics_constraints.json # Immutable IEEE-14, CCPP, and BESS physical bounds
β βββ agent_prompts_and_schemas.json # Schema-locked Jev System 1 & Bedrock System 2 contracts
βββ src/heliogrid/
β βββ config.py # Runtime BYOK key manager & resource loaders
β βββ data/ # ETT, CCPP, and IEEE 14-Bus Newton-Raphson AC power flow twin
β βββ ml/ # Psychrometric/STL features, BoostedTrees + PyTorch + Registry
β βββ optimization/ # SLSQP CCPP cooling/vacuum optimizer & constrained AC-OPF + BESS
β βββ decision/ # TypeSafe Jev (`typesafe/jev-1.13`) Noul/Choice/Score engine
β βββ autoresearch/ # Karpathy self-recurring loop & persistent experiment ledger
β βββ agents/ # LangGraph 7-node StateGraph, LangChain @tools & Amazon Bedrock
β βββ api/ # FastAPI server & Dark-Obsidian Double-Bezel Control Room UI
βββ scripts/
β βββ start_dashboard.py # CLI & BYOK Control Room launcher (--open-browser, --jev-key, etc.)
β βββ train_models.py # Full 9,568-hour training & .pt checkpoint serialization pipeline
β βββ verify_bedrock_live.py # Live AWS Bedrock & Jev connectivity diagnostic tool
β βββ run_benchmarks.py # Reproducible 168-hour AC-OPF & Karpathy loop benchmark suite
β βββ run_autoresearch.py # Standalone CLI for Karpathy continual learning loop
βββ tests/ # 7 Phased pytest verification suites (20/20 passing)
βββ quickstart.sh # One-command setup, model verification & UI launcher
βββ Makefile # Developer & operator shortcuts (make open, make test, etc.)
βββ program.md # Karpathy autoresearch declarative specification
βββ AGENTS.md # Engineering & anti-hallucination rules
βββ pyproject.toml # Package metadata & dependencies
Run the complete 7-phase test suite (20/20 tests covering real datasets, Newton-Raphson AC power flow, hybrid ensembles, SLSQP AC-OPF, TypeSafe Jev System 1 primitives, Karpathy self-recurring loop monotonic improvement, LangGraph + FastAPI endpoints, and .pt checkpoint SHA-256 integrity) and ruff linter:
make test
make lintAuthor: Rishabh Patil (@MrRobotop)
Repository: https://github.com/MrRobotop/HelioGrid-AI
License: Released under the MIT License. Copyright Β© 2026 Rishabh Patil.