One line: GRIDRUSH is an agent that decides which compute loads to energize, with what firmness, and with what flexibility, so that MW that would sit in an interconnection queue for years start earning in months — without violating the local grid or the availability the customer bought.
Built in one day at SF x MIT Hack Day: Energy (AGI House SF, 2026-11-08), track set “grid-friendly compute” (tracks 1 + 2 + 3: where → how much → how flexible).
| What | Where |
|---|---|
| Overview deck — the whole story in ten slides | pitch/GRIDRUSH-overview-deck.pdf |
| Overview video — the three-minute walkthrough | docs/GRIDRUSH-overview.mp4 |
| One-page leave-behind (A4 landscape) | pitch/GRIDRUSH-one-pager.pdf |
| Project source pack — one self-contained document, 18 sections | PROJECT_SOURCE_PACK.md |
| Public site with all dashboards and a one-click download | https://dmitrypehovski.github.io/gridrush/ |
| 3-minute verbatim demo script + cue card + 16-question Q&A | pitch/demo-script.md |
A data center can be built in 2–3 years; grid interconnection historically takes longer. ERCOT tracked ~410 GW of large-load requests in March 2026, ~87% of them data centers, against ~7.5 GW approved to energize since January 2022. On stage, the utility (PG&E) and the track host argued the same point from opposite sides: would you connect faster if you accepted 98% availability and curtailed a few hours a year?
GRIDRUSH answers that question with a computed portfolio instead of an opinion.
For 40 candidate edge sites (0.5–10 MW) sharing 8 substations, over a 36-month horizon, the agent chooses per site:
| Decision | Options |
|---|---|
| connect or not | skip / firm / flex (non-firm, step-up) / flex + battery |
| how much | 0.5 … min(nameplate, available headroom) MW |
| availability commitment | 99.99% (0.88 h/yr curtailment, revenue premium ×1.25) · 99.9% (8.76 h/yr, ×1.00) · 99% (87.6 h/yr, ×0.80) |
| speed | firm waits 18–36 months · flexible connects in 5–10 · with an onsite battery 1–3 months later |
The trade is explicit and physical:
- firm MW is capped by the substation's tightest hour of the year → small and slow;
- flexible MW can be several times larger immediately, but the site must drop load in stress hours;
- the SLA decides how many stress hours you may drop, and the contract price pays for reliability;
- a battery covers the deficit in those hours, at a capex.
The agent’s plan is then checked hour by hour: the evaluator recomputes every substation’s load from every site on it and scores grid limits plus each site’s own SLA. One honest detail — flexible load cannot exceed headroom by construction (that is what flexible service means: the site is clipped to what the grid can carry), so for the flexible portfolio the agent actually chooses, the binding check is the per-site SLA, not the violation count. Violations would appear if firm capacity were oversubscribed; we report both numbers and we say which one binds.
| Strategy | MW online at month 12 | Contract-compliant MW·months | Net value |
|---|---|---|---|
| All firm, wait for the queue | 0 | 651 | $65.1M |
| Sign everything non-firm at max MW | 173 | 3384 gross, mostly non-compliant | −$449.5M |
| GRIDRUSH agent | 85.4 | 2490.6 | +$279.9M |
- 4.30× the net value of the best honest strategy, 3.82× the sellable compute
- 0 MWh of grid violations, every site meets its own SLA (no SLA breach energy)
- 32 of 40 sites energized — the agent says no to the rest, which is a decision, not a failure
- the naive “speed at any price” strategy is loss-making: speed without reliability is not a business
The interesting part is how it wins: it does not pick one availability level globally. It pays for 99.99% where stress hours are rare (and collects the ×1.25 premium) on 16 sites, 99.9% on 6, and accepts 99% on 10 where the grid is congested — buying 5–10 months of earlier revenue. Batteries and firm service turn out to be dominated: the exact per-site trade makes them unnecessary.
The event asked for “agents that do that testing and get better at it each run”. GRIDRUSH is that loop, made explicit:
proposer → portfolio → simulator (hour-by-hour) → score → keep if better → mutate
loop/search.py— a fast mutator over a 9-parameter policy (≈200 generations in ~20 s), with an adaptive step and restarts; every generation is logged (results/loop.json → iterations) so the learning curve is real, not an illustration;loop/local_search.py— second proposer: per-site proposals (mode × MW × SLA × battery) scored by the full-year simulator; only SLA-clean improvements are accepted;rsi_agent/— the same loop with an LLM proposer: the agent writes new policy code (gen_0.py…gen_10.py), the simulator scores it, the best is kept (rsi_log.json,BEST.md). The archive survives between rounds: round 2 resumed fromgen_6and produced four more generations (rsi_log2.json,BEST2.md) — it converged at +0.12%, which is the honest finding (exact per-substation DP leaves little room), and one generation regressed and was rejected. This is where the result came from: the fast search plateaus at 179 (score), then the LLM agent rewrites the policy — an exact per-substation DP — and jumps to 279.9, beating even a hand-written greedy reference ($250M). 7 generations, 1 of them a recorded failure (returned an empty portfolio, scored 115) — kept in the log on purpose;- the policy also repairs its own plan (
repair_portfolio) — plan, verify, shrink MW until every site's SLA holds; without this step the plan is optimistic and part of the MW is unsellable.
Open web/regions.html for all four with links to their dashboards.
| Region | 2025 peak | Net value ratio | Compliant MW·months ratio | Violations |
|---|---|---|---|---|
| CAL (CAISO) | 43,860 MW | 4.30× | 3.82× | 0 |
| TEX (ERCOT) | 83,597 MW | 3.41× | 2.79× | 0 |
| MIDA (PJM) | 160,560 MW | 4.07× | 3.57× | 0 |
| MIDW (MISO) | 120,343 MW | 3.88× | 3.33× | 0 |
The main engine is an hourly capacity verifier on a headroom profile (the shape of data a utility
publishes — PG&E ICA gives 576 hourly values per line section). On top of it, sim/powerflow.py runs
a real AC power flow (pandapower mv_oberrhein: a real 20 kV German MV feeder, 179 buses,
2×25 MVA). Base case converges at min V 0.9756 p.u.; the feeder has 8.2 MW of spare capacity with
0 violations, while our five largest sites need 31.8 MW — they do not fit on one feeder (71
violations, voltage collapse). That is exactly why this is a siting problem: 0.5–10 MW edge sites
are the right unit, and the agent has to distribute them. web/physics.html shows every number from a measured run.
Two limits from one real feeder — 8.2 MW and 6.53 MW. Both are measured, and they answer different questions. 8.2 MW is the first thermal violation when the top-5 site profile is scaled uniformly (transformer T142 reaches 100.4 %). 6.53 MW is the conservative admission limit we enforce: the same total added load spread over the five best-headroom buses in proportion to their headroom, with every bus at or above 0.95785 p.u. and every branch at or below 100 %. We quote 6.53 on stage because it is the number the optimiser actually uses.
Does 4.30× survive if the economics are wrong? The ratio is invariant to the margin level —
the same prices scale the agent and the baselines together: 4.30× at $0.033, $0.10 and $0.30 per
MW-month (results/sensitivity.json). What is sensitive is the SLA premium spread: flattening
1.25 / 1.00 / 0.80 to 1.10 / 1.00 / 0.95, and halving the penalty, still gives 4.04×. The
physical ratio — 3.82× contract-compliant MW-months, 2490.6 vs 651.3 — contains no prices at all.
What rests on assumptions is the absolute dollar level ($279.9M), not the comparison.
Real PG&E data. sim/real_sites.py rebuilds the site model from the public PG&E GRIP Data
Portal (no key): 8 real substations with real coordinates, facility ratings and 2026 loadings.
Running the same agent on them gives 1.22× (397.7 MW online at month 12, 11,533 sellable
MW·months, zero violations) — the premium collapses because, at the substation aggregate level,
these substations are large and not the bottleneck. PG&E's own Grid Needs Assessment shows where the
constraint actually lives: in a 2,000-circuit sample, 124 of the 1,462 circuits that actually carry data
(8.5%; 538 of the 2,000 sampled rows are a −99.99 sentinel) are already deficient for 2026 (up to
12.4 MW). That is the next data we need, and it is why the ask is circuit-level ICA.
web/dashboard_real.html, results/loop_real.json, data/sites_CAL_real.json.
Physics in the loop. The measured admission limit — 6.53 MW of new load per MV feeder
(first violation at 7.03 MW) — is not just a check: it is passed into the portfolio builder
itself (build_portfolio(..., sub_limit_mw=...), driven by sim/physics_gate.py →
loop/run_physics.py), so the policy physically cannot over-subscribe a feeder; the final plan is
then verified with eight real AC-PF runs. With real physics inside the loop the same agent reaches
2.39× instead of 4.30× (net $155.5M): a real feeder takes ~6.5 MW, not the 20–60 MW the
synthetic capacities assumed. Both numbers are reported — the delta is the value of real utility
ICA data. See results/FINAL_NUMBERS.md.
Honesty is part of the demo. Every number on screen is labeled.
Real (public):
- hourly 2025 demand, day-ahead forecast and generation-by-fuel for the 12 largest US balancing
authorities — EIA-930, public domain (
data/us_grid_hourly_2025.csv, shipped in this repo); - the grid stress shape (
u[h] = demand / p95(demand)) that drives headroom comes from that data; - interconnection timelines order-of-magnitude, curtailment economics and the availability/premium logic are taken from the hackathon repo’s catalogue (MISO 18–36 month large-load process, Firm Service Step-Up, ERCOT batch studies) and from the talks on stage.
Synthetic (documented in sim/assumptions.md):
- 40 sites / 8 substations geometry, capacity, loading factors;
- exact queue months per site, battery capex ($0.35M/MW + $0.25M/MWh), revenue ($0.10M per MW-month), SLA price multipliers (1.25 / 1.00 / 0.80), SLA penalty multiplier;
- the 2025 year is replayed as a representative year across the 36-month horizon.
Data-quality caveats we found and handled: 7 corrupt demand rows in the CSV (interpolated), 26 rows
with an empty demand value (interpolated), one 24-hour ERCOT gap, and a physically implausible
solar_mw column (not used for capacity or stress
calculations; only the demand shape is used).
cd gridrush
python3 loop/run.py --region CAL --iters 180 # agent loop → results/loop.json
python3 web/build.py results/loop.json # single-file dashboard → web/dashboard.html
open web/dashboard.html # works offline, no network, no CDNDev/mock mode without the simulator: python3 loop/run.py --mock --iters 120.
| Path | What |
|---|---|
sim/ |
model: curves.py (single source of truth for headroom), engine.py (hour-by-hour verifier + economics), baselines.py (3 honest strategies), run.py → results/sim_baseline.json, assumptions.md |
loop/ |
RSI loop: policy.py (parameterized policy + self-repair), search.py (mutator), local_search.py (per-site proposer), run.py (CLI → results/loop.json) |
results/FINAL_NUMBERS.md |
the frozen numbers every pitch artifact references |
rsi_agent/ |
LLM-in-the-loop generations: eval.py, gen_*.py, rsi_log.json, BEST.md |
web/ |
dashboard.html — one self-contained offline file (loop replay, portfolio map, stress day, strategy comparison); regions.html — four operators + physics variant; dashboard_physics.html — physics-in-the-loop run; live_server.py — live run streamed to the browser (python3 web/live_server.py, ~11 s) |
pitch/one-pager.html + GRIDRUSH-one-pager.pdf |
one-page leave-behind for judges, generated from the frozen JSON by build_onepager.py |
pitch/ |
submission.md, demo-script.md (3 min, timed), qa-defense.md, hooks.md |
data/ |
EIA-930 hourly 2025 starter dataset (public domain) |
CONTRACT.md |
frozen model/API/JSON contract used to build this in 4 hours |
- Replace synthetic substation data with real utility hosting capacity (PG&E GRIP/ICA publishes 576 hourly values per line section) and real interconnection queue positions (MISO, ERCOT batch data).
- Replace the headroom model with AC power flow + N-1 (
pandapower/powermodelson synthetic grids) so capacity is a physics result, not a profile input. - Let the agent negotiate the contract, not just pick from it: durations, offtaker availability tiers, and multi-site aggregation into a single grid-service offer (track 4).
- Turn the loop into an operations loop: re-run weekly as queue positions and load forecasts move.
Ask: real hosting-capacity and queue data from a utility or an edge developer — one region is enough to replace the biggest synthetic assumption.