A Hume Fogg senior taught an AI to watch Nashville’s water 24/7 — Drinky for drinking water, Stinky for wastewater. Not a chatbot. A monitoring system that catches anomalies hours before human operators would, then hands the decision back to the operators who’ve done the work for decades. From the muddy Cumberland to a scholarship stage in Atlanta, all of it built on one rule her grandfather left her.
Mira Bowles had seen the Cumberland River her entire life — brown, churning, perpetually muddy — but she’d never really thought about it until the day her grandfather explained what he’d spent thirty years doing. They were on the pedestrian bridge downtown, a Tuesday lunch tradition they’d kept since she started at Hume Fogg. Hector held up a bottle of Nashville tap water, crystal clear in the afternoon sun.
How? Intake screening, coagulation, flocculation, sedimentation, filtration, disinfection. The short version, Hector said: they test it. Constantly. Every stage, every day, multiple times a day — pH, turbidity, chlorine, bacteria counts. The grin faded when Mira asked what happens if you miss something. “That’s the nightmare every water operator has. It’s happened other places. Flint. Camp Lejeune. So we test. We stay vigilant. We don’t miss.”
“I want to help with that,” Mira said. Not water engineering, she clarified. What if there was a way to test constantly — every second, with something that never gets tired or distracted or forgets to log a result? “You’re talking about automation.” “I’m talking about AI.” Hector pointed out the 8th Avenue Reservoir — built in 1889, the same year the Eiffel Tower opened, still holding 51 million gallons behind sound limestone walls 133 years later. Water systems aren’t just technology. They’re infrastructure, human judgment, institutional knowledge. “I’m not saying replace people,” she answered. “I’m saying help them.”
He smiled. “First lesson: the Cumberland River is the muddiest damn river in Tennessee, and making it drinkable is the hardest job in the state. If you can build something that helps with that, you’ll have my attention.” They walked back across the bridge, and Mira looked at the river differently now. Not just brown water. A problem to solve.
The first thing Mira learned was that she knew absolutely nothing about water treatment. The second was that her grandfather had forgotten more about it than most engineers would ever know. So she read — EPA guidelines, TDEC regulations, Metro Water annual reports back to 1993. She learned about Maximum Contaminant Levels and Total Trihalomethanes and the difference between turbidity and actual contamination. She learned the Cumberland wasn’t just muddy — it was dramatically muddy, the kind that had frustrated Nashville’s water supply since 1826.
She also learned that despite all the sophistication — the K.R. Harrington plant in Donelson processing 120 million gallons a day, every stage monitored and logged and reported to the EPA — water testing was still fundamentally manual. A human collected samples at scheduled intervals. It worked. Nashville had safe water. But it wasn’t optimized.
She wanted continuous intelligent monitoring. Sensors already existed — pH probes, turbidity meters, chlorine analyzers — but they just report numbers. They don’t understand context, don’t recognize patterns, don’t know that a 0.2 pH drop in thirty minutes is normal during high river flow but abnormal during low flow. That’s the AI part. Train a model on every test Metro Water ever logged. Teach it what normal looks like across seasons, river levels, treatment stages. Then flag deviations — before they become a problem. Mrs. Hooper called it ambitious. Mira knew. “Alright then. Let’s figure out how you’re going to do this.”
It was November 30th, 2022, when Mira first tried GPTP. She’d seen the headlines — Sam’s Place launches AI chatbot, people having philosophical conversations with it, concerns about academic integrity — and she’d been curious but skeptical. AI was mostly just better autocomplete, right? Pattern matching and statistical prediction?
Then she prompted it: “Explain the stages of drinking water treatment and the most common points of failure at each stage.” What came back was structured, detailed, accurate — not perfect, she caught a few oversimplifications, but far better than expected. She tried another about filter degradation in rapid sand filtration. Again: structured, specific, useful. She pushed into the relationship between river turbidity and coagulant dosing, seasonal variations in chlorine demand, coliform detection versus coliform risk. And the AI kept up.
It wasn’t thinking — Mira understood that. It was pattern-matching against vast training data. But that’s exactly what she needed. Pattern-matching. Anomaly detection. Finding signals in noise. If GPTP could hold a coherent conversation about water treatment, a specialized AI could absolutely monitor water quality in real time. Not a chatbot — a monitoring system. Sensor data feeding a machine learning model trained on decades of historical data. Not replacing human operators. Giving them superhuman awareness.
She opened a new document and started writing. Project Proposal: AI-Augmented Water Quality Monitoring. Working Title: Drinky and Stinky. (Hey, if she was going to spend the next year on this, it should at least have a name that made her smile.)
The hardest part wasn’t the coding — Mira was competent with Python, dangerous enough with TensorFlow. The hardest part was getting real water quality data. Not the sanitized annual reports. The raw hourly numbers with all their messy reality. Metro Water’s Public Information Office didn’t respond, then cited security concerns, then quoted a $4,500 processing fee for an “overly broad” records request. She didn’t have $4,500.
Her grandfather had an idea: go smaller. Cookeville — Tennessee Tech’s town, maybe 30,000 people, a much simpler system. Positioned as a student research project with the Tennessee Tech engineering department, and with a few calls from Hector’s conference contacts, Mira had a meeting scheduled for January 12th, 2023. Two days after that, she had five years of hourly water quality data.
The data was beautiful — not because it was clean (it was messy as hell, full of gaps and sensor errors and obvious typos where someone logged “6.8” as “68” in the pH column) but because it was real. She could see the patterns: pH rising as automated chlorine dosing kicked in, turbidity spiking after rains, chlorine residual dropping in summer. And the anomalies — a filter taken offline, a miscalibrated turbidity sensor reading exactly 0.7 NTU high, an impossible pH of 4.2 before someone caught the typo. She built Drinky in four layers:
Layer 1 — Data cleaning: flag impossible and statistically improbable readings, but don’t delete them, because sometimes impossible readings are the first sign of a real problem. Layer 2 — Baseline modeling: a dynamic range of “normal” for each parameter that understood time of day, season, river conditions, recent treatment history. Layer 3 — Anomaly detection: flag not just “outside range” but the patterns that historically preceded problems. Layer 4 — Human interface: send useful alerts, with context and a recommendation, not just a siren.
Running on Cookeville’s history, Drinky flagged a slow pH drift on June 18th, 2021 — every reading still inside the acceptable 6.5–8.5 range, no alarms, no operator concern. Then the real log showed pH falling to 6.3 hours later, a manual alert, a chemical feed pump malfunction. Drinky would have caught it three hours before the pH actually dropped out of range. She screenshotted the result and emailed her Cookeville contact. The reply, forty minutes later: “Holy shit. Can you test this on real-time data?”
Real-time testing meant Mira sitting in Cookeville’s control room on a Saturday in March, laptop wired into their SCADA system, feeding live sensor data into Drinky. The plant operator, Bill, fifteen years in, was skeptical but willing. “I’ve seen a lot of fancy monitoring systems. Most end up gathering dust.” “This one’s named Drinky.” Bill laughed. “Alright, points for honesty.”
For two hours: nothing. Green lights across the board. Then at 10:47 AM, Drinky flagged turbidity at Filter Bank 2 increasing faster than the influent increase would predict — 0.6 NTU against an expected 0.4, possible media compaction. Bill checked: still well within range, but a little higher than usual for similar river conditions. “I wouldn’t have caught that without checking. And I wouldn’t have thought to check because everything’s in the green.” They ran a backwash — not scheduled until Monday — and turbidity dropped to 0.3 NTU.
Three more Saturdays. Six anomalies flagged: four real issues caught early (filter degradation, chlorine feeder drift, flow sensor calibration error, pH trending low), two false positives. Four out of six — 67% precision. Not perfect, but useful. As Bill put it, even the false positives flagged things worth checking. “Kid, if you can make this work on the Cumberland River, you can make it work anywhere. I’ll write you a reference. And I’ll tell my boss this should be in our budget next year.”
So Mira drafted the email she’d been planning since September — to Metro Water Services’ Environmental Compliance Department, requesting a beta-testing partnership. She opened with her grandfather: thirty years turning muddy Cumberland water into clean drinking water for a million people, and the lesson that clean water isn’t magic — it’s vigilance. She attached the Cookeville results, Bill’s reference, the full technical documentation, and hovered over Send for thirty seconds before her grandfather’s voice echoed: “You can’t make it drinkable if you’re afraid to touch the river.” She clicked Send.
The response came two weeks later — not a form rejection but a personal email from Janet Reeves, Environmental Compliance Manager. The Metro Water headquarters on Lebanon Pike looked like it had been built in the 1970s and updated approximately never. Janet had a graying ponytail and a handshake that meant business. “You’re Hector’s granddaughter. I worked with him for six years. Good man. Careful operator. Never cut corners.” The conference room held four more people: Tom Hendricks (Plant Manager, K.R. Harrington), Sarah Hooper (Senior Operator, Omohundro), David Park (IT Systems Manager), Maria Santiago (Quality Assurance Lead).
Mira walked them through the architecture, the data cleaning, the baseline modeling, the June 2021 catch. Tom Hendricks stopped her: how did it catch a drift that was technically within normal range? “Because ‘within normal range’ isn’t the same as ‘normal.’ Your acceptable pH range is 6.5 to 8.5. But your actual pH is almost always 7.0 to 7.4. The AI learns what your normal actually looks like, not just what the regulations allow.” Sarah Hooper pressed on false positives. Cookeville showed 67% precision — and even the misses were worth checking, because the system explains why it’s flagging, with historical context. Human in the loop. Always. The AI provides information; operators decide.
Maria Santiago asked about EPA compliance — documentation gaps would be a non-starter. Mira was ready: the system logs everything, creating a permanent, time-stamped record more comprehensive than manual logs, and it supplements required sampling rather than replacing it. Silence. Then Janet asked what she was calling it.
Janet looked at the other four. Some silent communication passed. “Alright. Here’s the deal. We’ll pilot Drinky at Omohundro for three months. That’s our historic plant — running since 1889. If it can handle the Cumberland’s chaos there, it can handle anything. Read-only access to our SCADA data, weekly reports, and if your system ever interferes with operations or creates compliance concerns, we pull the plug immediately.” And one more thing: “Your grandfather would be proud of this. Don’t prove us wrong.” Mira felt something tighten in her chest. “I won’t.”
The Omohundro Water Treatment Plant sat on the Cumberland east of downtown — a Victorian-era pumping station expanded and modernized over 134 years but never fully replaced. The original 1889 pump house was still there: red brick, arched windows, industrial architecture that looked like someone decided water infrastructure deserved to be beautiful. Sarah Hooper gave the tour. “This is where your grandfather worked for his first ten years. Before they built Harrington, this was the only drinking water plant for the whole city.” Forty million gallons a day, still running, still backing up Harrington.
Sarah walked through the process — intake screens, coagulation, flocculation, sedimentation, rapid sand filtration, chlorination, pH adjustment — every stage tested hourly through the day. It worked. Nashville had great water. “But the Cumberland is temperamental. We’re good at responding to changes. But we’re reactive. What you’re proposing is predictive. Then let’s see if Drinky can handle the meanest river in Tennessee.”
It took two weeks for Drinky to calibrate. Then, June 7th, 6:23 AM, the first real alert: chlorine residual declining faster than the decay model predicted — 1.8 mg/L against an expected 2.0. Sarah checked the river and called back: algae, more than usual for early June, consuming chlorine. The system was working fine; they just needed to bump the dose. “We would have caught it this afternoon when residual dropped toward 1.0 and triggered our low-level alarm. You caught it at 1.8. That’s six hours earlier. Good catch, Drinky.”
Two weeks later, a severe thunderstorm dumped three inches of rain in four hours. The Cumberland went from 15–20 NTU to over 200 in ninety minutes — brown water thick with runoff. Drinky lit up: rising source turbidity, accelerated filter loading, sedimentation efficiency dropping, floc carryover. Sarah called: “Your AI is going nuts. Is it panicking?” No — everything it flagged was accurate, and it caught one thing she wouldn’t have noticed for another hour: Filter 3’s headloss rising faster than the others, its media more compacted. Extra eyes. Extra brain. Doesn’t get tired. Six hours, forty-three alerts, thirty-eight validated, one false positive.
The hardest test wasn’t dramatic. Just a slow pH drift over six days — 7.2, perfectly acceptable, historically unusual for mid-July at Omohundro, where the five-year average ran 7.4 to 7.6. Sarah physically checked the limestone feeder, which meters limestone for pH adjustment and corrosion control. It was delivering 85% of target dose. If it had stayed miscalibrated another two weeks, pH could have dropped below 7.0 and started leaching lead from old pipes in Nashville’s older homes. Mira’s stomach flipped.
Once Drinky proved itself, the question became wastewater. Mira met Janet Reeves and a new cast: Robert Jackson (Central Wastewater Plant Manager), Lisa Tran (Process Engineer), Jimmy Webb (Environmental Compliance, wastewater side). Robert was blunt: “Drinking water is relatively consistent — it’s river water, you treat it, it’s done. Wastewater is chaos. Every toilet flush, every industrial discharge, every rainstorm changes what’s coming in. And we have to treat it all.” Which, Mira said, sounded like exactly the kind of problem pattern recognition could help with.
Same architecture as Drinky, adapted: influent monitoring (what comes in), process monitoring (aeration, clarifier settling, biosolids), effluent monitoring (what goes out to the river, meeting EPA discharge limits). The goal was to catch trends — gradually rising influent ammonia hinting at an upstream discharge, slowly declining clarifier settling. Lisa Tran nodded slowly: “We do a lot of that by feel. Experience. Knowing how the plant sounds and smells and behaves. You’re trying to quantify that?” “I’m trying to augment it. Your experience is irreplaceable. But you can’t watch every parameter 24/7. Stinky can.”
Jimmy Webb asked why she was calling it Stinky. “Because wastewater treatment is the single most important public health service that nobody wants to think about. So if calling it Stinky makes one person remember that wastewater treatment matters? Mission accomplished.” Robert Jackson laughed, head thrown back. “Alright. I like you. Let’s test Stinky at Central. Same deal as Omohundro — you mess up our EPA compliance, we’re done.”
Central Wastewater Treatment Plant processed 60 million gallons of Nashville’s sewage per day, and it smelled exactly like Mira expected. Lisa walked her past massive aeration tanks where billions of bacteria consumed organic matter, clarifiers, anaerobic digesters. “The challenge with wastewater is that we’re not just removing contaminants. We’re managing a living ecosystem. Those bacteria are our workforce. If the water gets too cold or too hot or too toxic, they fail. And when they fail, we discharge untreated sewage into the river.” The early warning for bacterial failure was mostly experience — watching dissolved oxygen, the foam, microscope samples. Part art, part science. “Can Stinky learn the art part?” “We’re about to find out.”
The overnight operator, Jerome, called Lisa. By the time she arrived, pH had dropped to 6.1 — acidic enough to start killing the bacterial culture. “This is an illegal discharge. Somebody dumped something. We need to find the source before it destroys our bacteria.” Working backward through the collection system, they found the culprit by 5:00 AM: a metal finishing facility dumping acidic rinsewater into the sewer instead of its mandated holding tank. They stopped the discharge, added lime, and saved the culture. Without that early alert, the acid would have sat in the aeration tanks for four hours. Days to recover. Possibly a forced bypass.
The second catch was subtler — aeration tank dissolved oxygen showing unusual diurnal variation, evening DO drifting just barely below target. Not a violation, just slightly off. Lisa pulled previous Octobers: DO trending lower, consistently, on track to fall below 2.0 mg/L by mid-November, which meant under-aeration, inefficient bacteria, accumulating organic matter, eventual discharge failure. They increased aeration 8%. “That’s the value. Catching the slow trends before they become problems. Stinky gave us a six-week head start.”
By November, Mira was three months into her freshman year at Tennessee Tech — civil engineering, environmental focus — and Drinky and Stinky were running at two Nashville facilities. Her environmental engineering professor, Dr. Sarah Martinez, pulled her aside. “Have you heard of the TRU Foundation scholarships? One Chain — the rapper — runs a foundation that funds skilled trades and technical apprenticeships. They give major scholarships to students doing innovative work that bridges technology and practical application. Your project is exactly what they look for.”
“I’m not really doing trades work though.” “You’re doing water infrastructure work. That’s about as practical as it gets. And the application specifically wants AI implementation that augments human workers rather than replacing them. Sound familiar?” The requirements were exacting — demonstrated AI solving a real-world problem, human-in-the-loop design, a working prototype, a clear HBCU-to-career pathway, and a hard rule: NO AI-GENERATED ESSAYS. Submit a live PromptDeck presentation, a professional research paper, or technical innovation documentation.
Mira spent December and Christmas break preparing. Janet Reeves approved presenting Metro Water data immediately — a win-win. Sarah Hooper and Lisa Tran got her video footage from Omohundro and Central. She prepared three physical samples — Cumberland River water (brown, turbid), mid-treatment (clearer, unfinished), finished drinking water (crystal clear) — and built a live demo: a portable turbidity sensor wired to her laptop, running Drinky in real time. The presentation was set for January 14th, 2024, at TRU Foundation headquarters in Atlanta. She practiced until she could do it in her sleep.
The scholarship board room held seven people: One Chain (ONEED PEEPS), foundation founder; Dr. Marcus Williams (Morehouse, HBCU education pipeline); Sharon Mitchell (workforce development); Carlos Hernandez (technology innovation); Dr. Janet Foster (environmental engineering, Georgia Tech); Marcus Henderson (HVAC precision training, former Blue Angels pilot); and Reverend Thompson (faith community partnerships). Mira set up her three beakers at the front. “My name is Mira Bowles. I’m a freshman at Tennessee Tech studying civil engineering. And for the past year, I’ve been teaching AI to understand water.”
She held up the Cumberland sample — brown, muddy, full of sediment and algae and fish waste, completely normal for one of the muddiest rivers in Tennessee. Then the finished drinking water — crystal clear, safe, rigorously tested. “My grandfather, Hector Bowles, spent thirty years at Metro Water Services making that transformation happen every single day.” The question she’d asked: what if AI could watch every stage 24/7 and flag anomalies before operators would even know to look? She walked them through the June pH incident, the July limestone feeder drift, the summer storm.
Then the live demonstration. She tested all three samples with the portable turbidity sensor. Cumberland: 187 NTU, normal for winter river conditions. Mid-treatment: 4.2 NTU, slightly high — recommend filter check. Finished: 0.08 NTU, well below the EPA limit of 0.3. “That mid-treatment reading is technically fine. But Drinky flags it because historical data shows that when post-clarification turbidity is above 4.0, there’s a 23% chance of filter loading issues within six hours. That’s not replacing human judgment. That’s giving operators information — six hours of warning instead of reacting after filters clog.”
Dr. Janet Foster asked about false positives: 11.6% over three months at Omohundro, 8.3% over two months at Central — and most were technically true, the pattern unusual even when no problem followed. Carlos Hernandez asked about scaling: the core architecture is transferable, but every system needs training on local data. Then Marcus Henderson, the former Blue Angels pilot, leaned in: what did her grandfather think of the project?
The room went quiet. Then One Chain spoke for the first time: “Tell us about Stinky.” Mira laughed, breaking the tension, and walked them through the October industrial discharge, the DO trend, the biological process monitoring. “Wastewater treatment is invisible infrastructure. But that waste is getting treated by people protecting everyone downstream. And they deserve the best tools we can give them.” Asked about her career plan, she said: deploy Drinky and Stinky at scale, then work for EPA setting standards for AI-augmented monitoring — open-source, adaptable, accessible. Asked why it mattered personally, beyond legacy: “Because water is dignity. That’s not luxury infrastructure. That’s basic human dignity.”
One Chain looked at the board, then back at Mira. “We’re gonna need a few minutes to deliberate. Can you step outside?” Five minutes felt like five hours. Then the door opened. The board was standing — all seven of them.
Mira’s hands started shaking. She shook his hand, tried to speak, finally managed: “Thank you. I won’t let you down.” “We know you won’t. Because your grandfather trained you right. And because you’re doing this for the right reasons.”
The scholarship changed everything — not because of the money, but the validation. When TRU Foundation announced the recipients in February 2024, Drinky and Stinky made national news: Water Quality Weekly (“Teen’s AI System Catches Water Treatment Issues Before Human Operators”), Civil Engineering Magazine, Environmental Protection. Beta-testing requests started flooding in.
Janet Reeves called in March: full deployment at Omohundro and Central, plus pilots at K.R. Harrington (drinking) and Dry Creek (wastewater). “We’ll provide two staff engineers. This is becoming operational infrastructure, not a student project. You’ll still lead the AI development, but we’re building this into our standard monitoring systems.” By April, the system ran at four facilities.
The Metro Water board approved permanent deployment at all four facilities. And three other Tennessee cities requested pilot programs.
Mira sat in the front row of Gateway Center Arena in a navy blazer her mom had bought specifically for this, trying not to be nervous. Ten thousand people filled the arena for the final day of awards. One Chain was receiving the Andrew J. Young Humanitarian Award for TRU Foundation’s 847 apprenticeships. PYELER TERRY was giving the keynote. And three students were being honored for “AI innovation that augments rather than replaces human capability”: Mira Bowles (Nashville, Drinky and Stinky); James Park (Portland, the Lighthouse Bias research); Kai Martinez (San Jose, the Patience Problem in conversational AI).
Ambassador Andrew Young himself presented the award — ninety years old, stooped but steady, voice still strong. “ONEED PEEPS understands what Dr. King taught us. That all labor that uplifts humanity has dignity. Through TRU Foundation, he’s proven that investment in people — real investment, not charity — creates lasting change.” One Chain accepted with visible emotion: “I stand here because people invested in me. And now I have the opportunity to invest in others. 847 apprenticeships aren’t just numbers — they’re 847 people building careers, supporting families, strengthening communities.”
Then he called the three students to the stage. Mira walked up on legs that felt made of water. “Mira,” One Chain said, “tell everyone what Drinky and Stinky do.”
The arena erupted. James explained his Lighthouse Bias work — hidden biases in AI training data. Kai explained the Patience Problem — AI that can’t tell when humans are done talking, learning that sometimes silence means thinking, not your turn. One Chain handed each of them a certificate: full scholarship, tool funding, HBCU pipeline connection, recognition as TRU Foundation AI Innovation Fellows. “Technology doesn’t have to destroy jobs. When designed ethically, when implemented with respect for human expertise, technology creates opportunities.” Ambassador Young stepped forward one more time: “Dr. King used to say that the arc of the moral universe is long, but it bends toward justice. You three are bending that arc. Keep going.”
By July 2025, Drinky and Stinky were running at eleven facilities across Tennessee — all five Nashville drinking water plants and all three wastewater plants, Cookeville fully deployed, Chattanooga in beta, Knoxville requesting a pilot. The EPA had issued a preliminary guidance document, “AI-Augmented Water Quality Monitoring: Best Practices and Human-in-the-Loop Requirements” — Mira’s work cited seven times. Tennessee Tech had added a course, “Machine Learning for Water Infrastructure,” and Mira was the teaching assistant.
Then, on a Tuesday in July, Janet Reeves called about the 8th Avenue Reservoir — the historic 1889 structure. A major renovation was coming, and the Metro Council had specifically requested that Drinky be part of the monitoring package. “They want AI protecting the city’s oldest piece of water infrastructure.” Mira’s chest tightened: her grandfather had helped maintain that reservoir for twenty years.
Mira managed to say thank you, ended the call, and sat for a few minutes in her dorm room, thinking about muddy river water becoming crystal clear. Thinking about thirty years of vigilance. About AI that augmented rather than replaced. About her grandfather’s voice: You can’t make it drinkable if you’re afraid to touch the river. She’d touched the river. Built something that worked. Shared it with the world. And now that muddy Cumberland water — brown and chaotic and fundamentally undignified — was being transformed into clean drinking water by skilled human operators using tools she’d built. Technology and humanity. Strengthening each other.
Same river, same city
The methodology