Madhvansh Choksi — building production AI systems for industrial water treatment.

I don't make demos. I make systems that run 24/7 on real cooling towers, replacing chemical dosing decisions that used to require a human operator staring at a SCADA screen.

Currently shipping TGF — an autonomous cooling tower control system that predicts water chemistry 6–24 hours ahead and optimizes chemical dosing across 7 simultaneous treatment chemicals using Model Predictive Control.

The short version: Nalco and ChemTreat charge $50K+/year to reactively dose 1–2 chemicals. TGF predicts the future and optimizes all 7 at once — delivering 15–30% chemical savings with zero critical failures.

Industrial cooling towers waste billions of dollars annually on chemical treatment because dosing is reactive — operators wait for pH to drift, then dump chemicals. Meanwhile, scaling corrodes heat exchangers, biofouling clogs fills, and plants lose efficiency.

Every major vendor (Nalco/Ecolab, ChemTreat, Solenis) sells the same thing: a fluorescent tracer that measures one chemical accurately, then charges you for the privilege of vendor lock-in.

Sensors (pH, Conductivity, Temp, ORP)
        │
        ▼
┌───────────────────────┐     ┌────────────────────────────┐
│  Chronos-2 Forecaster │────▶│  Statistical Fallback      │
│  (Zero-shot p10/50/90)│     │  (when GPU unavailable)    │
└──────────┬────────────┘     └────────────────────────────┘
           │
           ▼
┌───────────────────────┐     ┌────────────────────────────┐
│  Physics Engine       │◀───▶│  Chemical Residual Tracker  │
│  (LSI/RSI/CoC/Risk)   │     │  (Mass balance × 7 chems)  │
└──────────┬────────────┘     └───────────┬────────────────┘
           │                              │
           ▼                              ▼
┌──────────────────────────────────────────────┐
│  MPC Dosing Optimizer                         │
│  scipy L-BFGS-B · 2-hour receding horizon    │
│  Cost = chemical_cost + risk_penalties        │
└─────────────────┬────────────────────────────┘
                  │
                  ▼
┌──────────────────────────────────────────────┐
│  Safety Layer                                 │
│  Sensor fault detection · Hard limits         │
│  Rate limiting · PID backup · Emergency stop  │
└─────────────────┬────────────────────────────┘
                  │
                  ▼
           Pump Commands + Blowdown

• MPC over RL (SAC/PPO): Hard safety constraints are guaranteed, not learned. Works with 5K samples. Explainable to plant operators.
• Chronos-2 over PatchTST: Zero-shot works immediately on new towers. PatchTST needs fine-tuning on data we don't have yet.
• Mass balance over virtual sensors: We tried ML-based virtual sensors for hardness/alkalinity — R²=0.37 was too unreliable. Physics-based mass balance with weekly lab calibration is honest engineering.
• Statistical fallback always ready: Every Chronos-2 call has a Holt-Winters backup. No single point of failure.

tech_stack = {
    "ai_ml": ["PyTorch", "Chronos-2", "MOMENT", "TransNAS", "scikit-learn"],
    "optimization": ["scipy (L-BFGS-B)", "Model Predictive Control"],
    "backend": ["FastAPI", "SQLite (WAL)", "uvicorn"],
    "physics": ["Langelier SI", "Ryznar SI", "Arrhenius decay", "Mass balance"],
    "infra": ["Real-time dashboards", "Alert systems", "Sensor simulation"],
    "research": ["State Space Models (S4/Mamba)", "CoreML", "Time-series AD"],
    "languages": ["Python", "JavaScript", "Java", "SQL", "HTML/CSS"],
}

• MOMENT foundation model integration — reconstruction-based anomaly detection plugged into the TGF control loop (architecture is ready, awaiting model fine-tuning)
• Multi-tower orchestration — coordinated dosing across cooling tower farms sharing blowdown/makeup water
• Edge deployment — running the full MPC stack on industrial edge hardware (Raspberry Pi + sensor HATs for proof-of-concept)

I use Claude extensively in my development workflow — from debugging MPC cost functions to exploring Chronos-2 integration patterns to writing the physics engine's Arrhenius decay model. The TGF codebase is deeply intertwined with Claude-assisted development.

What Claude Max would unlock:

• Faster iteration on the MOMENT anomaly detection integration
• Multi-file refactoring across the 20+ module codebase
• Exploring novel MPC formulations with longer planning horizons
• Writing comprehensive test suites for safety-critical dosing logic

This isn't a side project. It's production software targeting a $10B+ industrial water treatment market where the incumbents haven't innovated in decades. Claude is the only AI assistant that can reason about the physics and the code simultaneously.

Building the future of industrial AI — one control cycle at a time.

If cooling towers interest you (or if you think MPC is underrated), let's talk.