Singapore runs ninety traffic cameras around the clock. Some are 1080p. Eleven are 320×240 hardware from the early 2000s. They look out over monsoon storms, midnight glare, peak-hour chaos, and the dead silence at 4am. And every single frame — every camera, every condition — gets piped through the same generic YOLO that treats them all identically. Which is exactly when vehicles get missed.

I kept staring at this and thinking: the model already has access to everything it needs to do better. The camera knows which camera it is. The system has live weather. The clock is right there. The PM2.5 air-quality index is one public API call away. So why is a clear afternoon highway and a rain-soaked midnight feed processed the same way?

The answer turned out to live in a 2018 paper — FiLM (Feature-wise Linear Modulation, Perez et al.). Tiny modules that take an external signal and use it to scale and shift the network's internal features. Three reasons it was the right tool:

• Identity at initialization. γ=1, β=0 means day zero is exactly vanilla YOLO — the model only specializes where it helps loss. No regression risk.
• Tiny. ~130K parameters on YOLO's 9.4M. A 1.4% overhead with negligible inference cost.
• No retraining. One model, with a dial for context. Not ninety separate models. Not an ensemble.

Picture the vision network as layers of pattern detectors stacked on top of each other. With FiLM, I dial each detector's volume up or down based on the current weather, time, camera, and air quality. In monsoon rain, the detectors looking for crisp edges get turned down. The ones looking for blurry motion blobs get turned up. At 2am on Camera #47, the network behaves differently than at 2pm on Camera #12. The model figures out the dialing on its own — I just feed it the current state of the world.

The FiLM transform itself is two affine parameters per channel:

feature_out = γ(context) ⊙ feature_in + β(context)

A tiny encoder (MLP) processes weather code, temperature, sin/cos hour-of-day, a 16-dim camera ID embedding (learned per camera), resolution, and PM2.5, then predicts γ and β for the P3/P4/P5 stages of YOLOv11's backbone. Each of the 90 cameras gets its own learned embedding that captures viewpoint and quirks.

%%{init: {'theme':'dark', 'themeVariables': {'primaryColor':'#22d3ee','primaryBorderColor':'#22d3ee','lineColor':'#64ffda','primaryTextColor':'#e5e7eb','background':'#0d1117'}}}%%
flowchart LR
    CTX[live context<br/>weather · time · camera · PM2.5] --> ENC[MLP encoder<br/>γ, β]
    IMG[camera frame] --> BB[YOLOv11 backbone]
    BB --> F1[FiLM P3]
    BB --> F2[FiLM P4]
    BB --> F3[FiLM P5]
    ENC --> F1 & F2 & F3
    F1 & F2 & F3 --> H[detection head]
    H --> OUT[vehicles · plates · tracks]

The system is live on HuggingFace Spaces, polling all 90 cameras every 60 seconds. 80,000+ records and growing. Dataset publishing to Kaggle for open traffic research.

Every AI hyperscaler needs power — gigawatts of it, twenty-four seven. Connecting a new data center to the grid takes three to seven years. Your competitors aren't waiting. So you build your own natural-gas plant on-site, behind the meter. The catch is picking where, and the catch is brutal: it's a three-body problem. Land has its own constraints (zoning, fiber, water, flood). Gas supply has its own (pipeline reliability, hub distance, curtailment risk). Electricity markets have their own (LMP spreads, price regimes, scarcity events). Real consultants charge six figures and take months because they evaluate these axes one at a time.

I wanted the three to argue with each other in real time instead of sequentially. A great parcel with bad gas economics shouldn't beat a decent parcel with stellar gas economics, but spreadsheets can't tell you that. So I gave each axis its own ML model — picked deliberately, not by default:

• Land → Random Forest. Parcel features are tabular and noisy. RFs handle that gracefully and give per-site SHAP attributions, so I can explain why a site scored where it did.
• Gas → GPU Gaussian KDE. Pipeline incidents cluster spatially. KDE reads that distribution out cleanly without needing labels, and PHMSA's incident database is public and severity-tagged.
• Power → Random Forest + GMM. The market itself has distinct regimes — normal, wind-curtailment, scarcity. A 3-component GMM finds them in unlabeled ERCOT data. The RF then predicts spread durability conditioned on the regime.
• Composite → TOPSIS. A multi-criteria decision method from 1981 that's still the cleanest way to combine apples and oranges with user-adjustable weights.
• Agent → 7-node LangGraph + Claude. Five intents — stress-test, compare, timing, explanation, config — let a human drive the analysis without writing code. Routes through parse_intent → tool node → synthesize and streams back as SSE tokens.

Click any coordinate across Texas, New Mexico, or Arizona — ERCOT or WECC market — and the system pulls live market data, looks up the parcel, scores all three axes, blends them through TOPSIS, and runs 10,000 Monte Carlo simulations to give you P10/P50/P90 cost ranges over the next 20 years. A LangGraph agent on top of Claude reads all of that — plus 72-hour LMP forecasts from the Moirai foundation model — and answers your follow-up questions in plain English. "What if gas prices spike 30%?" re-runs the analysis on live ERCOT and CAISO feeds.

Wired to ten public data sources — PHMSA pipeline incidents, ERCOT and CAISO LMP, EIA gas prices, NOAA weather, PERM-A federal land ownership, FCC fiber maps, FEMA flood zones, GridStatus, and live Tavily web enrichment — refreshing every five minutes through APScheduler background jobs. Pandera schema validation gates every row; failures get quarantined, never silently dropped.

%%{init: {'theme':'dark', 'themeVariables': {'primaryColor':'#fb923c','primaryBorderColor':'#fb923c','lineColor':'#64ffda','primaryTextColor':'#e5e7eb','background':'#0d1117'}}}%%
flowchart LR
    SRC[10 public sources<br/>PHMSA · ERCOT · CAISO · EIA · GridStatus<br/>NOAA · PERM-A · FCC · FEMA · Tavily] --> ETL[DuckDB · Parquet · Pandera]
    ETL --> M1[Land<br/>Random Forest]
    ETL --> M2[Gas<br/>GPU KDE]
    ETL --> M3[Power<br/>RF + GMM]
    M1 & M2 & M3 --> TOP[TOPSIS]
    TOP --> NPV[Monte Carlo · 10K]
    NPV --> AGT[LangGraph · Claude]
    AGT --> UI[map · scorecard]

Background jobs: GridStatus + Waha + regime every 5 min, Tavily news every 30 min, Moirai forecast every hour. Live ERCOT LMP also pushes via WebSocket at /ws/lmp/stream.