Pet food bag feeding guidelines are formulated for intact (un-neutered), young, active dogs, but 85% of pet dogs in the U.S. are spayed or neutered and most live sedentary indoor lives, meaning the printed portions deliver 20-30% more calories than these animals need daily. This systematic overfeeding has driven pet obesity to 56% of dogs and 60% of cats, cutting 2+ years off lifespan and causing diabetes, joint disease, and heart failure. The problem persists because manufacturers have a financial incentive to recommend larger portions since it increases purchase frequency, there is no regulatory requirement to differentiate feeding guidelines by activity level or neuter status, and veterinarians see pets too infrequently to catch gradual weight gain before it becomes clinical obesity.
Real problems worth solving
Browse frustrations, pains, and gaps that founders could tackle.
The U.S. Department of State Office of Authentications is the sole authority for apostilling federal documents like FBI background checks, yet it operates as a single centralized office in Washington, D.C. Standard mail-in processing takes 6-8 weeks, and even hand-delivered requests take 2-3 weeks. This bottleneck directly blocks immigrants who need apostilled FBI checks for visa applications in other countries, often causing them to miss embassy appointment deadlines and restart the entire application cycle. International job offers fall through, university enrollment windows close, and people remain in legal limbo for months. The problem persists because the State Department has no mandate or budget to scale this office, and there is no federated model allowing regional offices or embassies abroad to issue apostilles on federal documents.
Self-storage operators use introductory rates to fill units, then impose 10-25% rent increases every 6-12 months, knowing that the cost and hassle of renting a truck, hiring labor, and moving everything to a competing facility exceeds the annual rent increase. Public Storage alone was sued in a class action for systematically targeting long-term tenants with above-market hikes because data showed fewer than 5% of renters move out after a price increase. This persists because only a handful of states (CT, NY) cap storage rent increases, the industry consolidated to where four REITs control over 25% of all U.S. facilities, and storage leases are month-to-month by design so operators can raise prices with just 30 days notice without violating any contract.
AAFCO regulations require pet food labels to list crude protein and fat as minimums and crude fiber and moisture as maximums, meaning a bag claiming 20% minimum protein could contain 30% and a bag claiming 4% maximum fiber could actually have 12-16% total dietary fiber. This matters because owners of dogs with kidney disease must restrict protein precisely to avoid renal damage, and owners managing diabetic cats need accurate carbohydrate data that cannot be derived when every number on the label is a floor or ceiling rather than an actual value. The system persists because requiring actual nutrient ranges would force manufacturers to tighten production tolerances across batches, increasing manufacturing cost, and the pet food industry successfully lobbied to keep the guaranteed analysis format unchanged since the 1970s.
Over 60% of interstate moves booked online are sold by brokers who buy Google Ads using carrier-like names such as 'National Van Lines Moving,' collect a non-refundable deposit, then subcontract the actual move to whichever carrier is cheapest and available, without disclosing the subcontracting until pickup day. The customer has no say in who actually handles their belongings, cannot vet the carrier's safety record, and has no contractual relationship with the company that shows up. This persists because FMCSA requires brokers to register but does not require them to disclose their broker status in advertising, Google does not verify moving company license types before selling ads, and the deposit is structured as non-refundable so by the time the customer discovers the bait-and-switch it is too late to switch without losing $500-$1,500.
Emergency room patients and their families arrive in crisis, park in the nearest hospital garage, and have zero ability to predict how long they will be inside. The average ER visit lasts 4.5 hours, and complex cases stretch to 8-12 hours, racking up $30-60 in parking fees while someone is getting stitched up or waiting for test results. Patients sometimes leave the ER mid-treatment to move their car or feed the meter, delaying care and risking discharge. Hospitals outsource parking to third-party operators like LAZ and SP+ under revenue-sharing contracts that incentivize maximum fee extraction, and these operators have no relationship with the clinical mission of the facility. Validation systems exist but require patients to remember to ask at discharge — something people in medical distress routinely forget.
A 2026 Clean Label Project study of 79 dog food products found dry kibble contains arsenic at 11.8x, lead at 21.2x, and mercury at 24x the levels in fresh/frozen dog food, with concentrations 3-13x higher than equivalent human foods. This matters because dogs eat the same food daily for years, causing heavy metals to accumulate in their kidneys and livers, contributing to chronic disease and cancer, which is the leading cause of death in dogs. The problem persists because the FDA's action level for lead in pet food is 10,000 ppb, orders of magnitude higher than the 50 ppb limit for bottled water, and there is no regulatory requirement for manufacturers to test finished products for heavy metal contamination before shipping.
Many parking garages display a green 'clearance OK' light at the entrance based on a single overhead sensor, but the actual clearance drops by 6-12 inches deeper inside due to hanging pipes, sprinkler heads, and sloped ramps between levels. Drivers of full-size trucks and SUVs enter confidently, then hear the sickening crunch of their roof rack or cargo box hitting infrastructure two turns in. Insurance typically denies these claims as driver negligence, and the garage operator's posted signage ('enter at your own risk') shields them from liability. This persists because garage operators measure clearance at the entrance beam only and have no financial incentive to map internal obstructions, while retrofitting variable-height sensors throughout the structure costs $50K+ that generates zero additional revenue.
AAFCO's ingredient definitions allow independent rendering plants to combine dead animals from farms, shelters, and roadways into a single batch and label the output as generic 'meat meal' or 'animal fat' with no species identification required. This matters because pet owners with dogs that have specific protein allergies cannot reliably avoid triggering ingredients, leading to chronic GI issues, skin reactions, and expensive veterinary elimination diets that cost $500-$1,500 to diagnose what should have been on the label. The system persists because the FDA refuses to require species-specific labeling for rendered ingredients, the rendering industry lobbies to keep definitions broad so they can process mixed-source material at lower cost, and AAFCO is a voluntary association with no enforcement authority.
Rogue moving companies load a customer's entire household onto the truck, then refuse to unload at the destination unless the customer pays $2,000-$5,000 more than the binding estimate in cash or money order on the spot. This matters because the customer's belongings are literally held captive: filing an FMCSA complaint takes 6-18 months to resolve, local police consider it a civil dispute and refuse to intervene, and the customer often has children starting school or a job starting Monday so they pay under duress. The practice persists because FMCSA licenses over 14,000 interstate movers but employs fewer than 100 investigators, penalties max out at $10,000 per violation which is less than the profit from a single hostage job, and unlicensed brokers who subcontract to these carriers face zero liability for the carrier's behavior.
When an earthquake, hurricane, or wildfire destroys local telecommunications infrastructure, search-and-rescue teams arrive to find zero cellular connectivity across the entire disaster zone — exactly when AI-powered coordination (triaging damage reports, routing rescue teams, translating for multilingual victims, summarizing field reports for incident command) would save the most lives. The 2024 Maui wildfire, 2023 Turkey earthquake, and 2024 Hurricane Helene all destroyed cellular infrastructure within the first hour, creating 6-48 hour communication blackouts during the most critical rescue window. Cloud-based AI tools that worked perfectly in the staging area become completely non-functional the moment teams enter the disaster zone. A mesh network of Raspberry Pis running Gemma 4 at each forward operating point can provide local AI capabilities — summarizing handwritten triage tags from field medics, translating between rescue teams from different countries, generating structured SITREP reports from voice dictation — all operating on battery or generator power with zero dependency on external infrastructure, at a cost of $50 per node versus $5,000+ for ruggedized satellite terminals.
Every voice command to Alexa, Google Home, or Siri sends a raw audio recording to cloud servers for transcription and intent classification -- and these recordings include background conversations, arguments, medical discussions, and children's voices that the user never intended to share. Amazon confirmed in 2019 that human reviewers listened to Alexa recordings, and class-action lawsuits over unauthorized voice data collection have reached $30M+ in settlements. Users who want smart home voice control currently must accept that every sound in their home is transmitted to and stored on corporate servers, because the speech-to-intent pipeline runs entirely in the cloud. A Gemma E4B model running on a local smart home hub processes wake-word detection, speech-to-text, and intent classification entirely on-device -- the audio literally never leaves the living room, making privacy violations architecturally impossible rather than just policy-promised.
Of the world's 7,000 languages, 3,000 are endangered with fewer than 10,000 living speakers, and these languages are functionally invisible to commercial AI — OpenAI, Google, and Anthropic will never build dedicated models for Warlpiri (3,000 speakers) or Navajo (170,000 speakers with complex verbal morphology that breaks standard tokenizers) because the economics do not justify it. Elder speakers die at a rate of one every two weeks in many communities, taking irreplaceable grammatical knowledge with them. A Raspberry Pi running a fine-tuned Gemma 4 model serves as a community-owned language preservation and teaching tool: elders record speech that trains the model locally, the model then powers interactive language lessons for young community members, generates new example sentences in the language, and provides real-time translation assistance — all running on hardware the community owns permanently with no recurring API costs, no data leaving tribal sovereignty, and no dependency on a corporation that could deprecate the service. Open-source fine-tuning is structurally necessary because no commercial entity will ever invest in training a Warlpiri language model.
Coordinated agricultural drone swarms spraying pesticide across large fields must make real-time collision avoidance and path-replanning decisions every 100ms, but at 50-100 meters altitude over open farmland, cellular connectivity drops unpredictably -- a 3-second cloud API timeout during a formation maneuver means two $5,000 drones collide and crash into the crop they were protecting. A swarm of 12 drones can cover in 20 minutes what takes a single drone 4 hours, but only if they can coordinate without connectivity gaps. Cloud-dependent swarm coordination fails because the connectivity is not just slow but intermittent -- the drone has signal, then does not, then does again -- and any gap during a coordinated turn is catastrophic. Each drone running an on-device Gemma model for local path planning and neighbor-awareness makes independent collision-avoidance decisions using onboard LIDAR and inter-drone mesh radio, so cloud dropout does not cause loss of coordination or aircraft.
A large automotive factory with 10,000 vibration, temperature, and acoustic sensors sampling at 1kHz generates 864 million inference requests per day for quality anomaly detection -- at $0.0001 per cloud API call (the cheapest tier), that is $86,400/day or $31.5M/year, which is economically absurd for detecting whether a weld sounds right. Even at aggressive volume discounts, the per-query cost model of cloud APIs breaks down completely when sensor counts reach industrial scale, because the pricing is designed for human-initiated queries (hundreds per user per day), not machine-initiated queries (millions per sensor array per day). The structural issue is that cloud API pricing is per-query while sensor data is continuous and high-frequency -- the economic models are fundamentally incompatible. Running Gemma on a $200 edge device at each sensor cluster eliminates per-query costs entirely, reducing the total cost of inference from millions per year to a one-time hardware investment of $50,000 for the entire factory.
Underground mines at depths of 300-1500 meters have zero radio signal penetration — no WiFi, no cellular, no satellite — yet miners work alone in confined spaces where hands-free voice interaction with equipment checklists, gas sensor readings, and emergency protocols would prevent the 15,000+ mining fatalities that occur globally each year. Existing cloud-based voice assistants (Alexa, Google Assistant, Siri) are fundamentally useless underground because they require constant internet connectivity for speech recognition. Wired communication systems exist but are fixed to specific locations, so a miner 200 meters from the nearest phone station who detects unusual sulfur smell cannot quickly cross-reference gas safety thresholds while simultaneously evacuating. A Raspberry Pi strapped to a miner's belt running Gemma 4 with a microphone can provide fully offline voice-activated safety assistance: the miner says 'hydrogen sulfide 15 ppm' and the device instantly responds with evacuation protocol, safe direction based on last known ventilation map, and time-to-dangerous-exposure — all without any network infrastructure.
Municipal water treatment plants operate on air-gapped SCADA networks that are physically disconnected from the internet to prevent cyberattacks -- after the 2021 Oldsmar, Florida incident where a hacker remotely increased sodium hydroxide to 100x safe levels, most utilities hardened their air gaps. This means cloud-based AI for anomaly detection (detecting poisoning, equipment malfunction, or chemical dosing errors) is architecturally impossible -- there is no network path for sensor data to reach a cloud API or for the API response to reach the SCADA system. The air gap exists for critical safety reasons and will never be removed, so the only path to AI-powered anomaly detection is running models on-premise, on the same isolated network. A fine-tuned Gemma model running on an edge server inside the air-gapped network continuously analyzes flow rates, chemical levels, and pump behavior to detect anomalies that rule-based SCADA alarms miss, without ever creating a network pathway that an attacker could exploit.
Small-fleet commercial fishing vessels (under 60 feet) operate 50-200 miles offshore for 2-3 weeks at a time where cellular connectivity does not exist, and satellite data via Iridium or Starlink Maritime costs $250-1,000/month — prohibitive for owner-operators already earning thin margins. These crews make dozens of daily decisions (where to set gear, when to pull nets, whether an approaching weather system requires running to port) based on gut instinct and decades-old rules of thumb rather than data analysis. A single bad weather decision kills an average of 42 commercial fishers per year in the US alone, making it the deadliest civilian occupation. A Raspberry Pi running Gemma 4 can ingest local barometric pressure readings, sea temperature, wind data from onboard instruments, and the captain's text-based observations to provide natural-language weather risk assessments, optimal routing suggestions, and catch prediction based on conditions — all processed locally with zero connectivity cost, fine-tuned on the specific fishery and vessel type rather than generic global weather models.
Soil moisture sensors across a 500-acre farm detect that Zone 14 is at wilting point and needs water within the hour, but the farm's field zones are 2+ miles from the nearest WiFi router and have spotty cellular at best -- a cloud API call to classify soil conditions and decide irrigation timing either fails entirely or takes 30+ seconds during peak tower congestion. Over-irrigating by even 15% wastes $8,000-$12,000 per season in water costs for a mid-size farm, and under-irrigating at a critical growth stage permanently reduces yield by 20-30%. Cloud models cannot solve this because the connectivity gap is geographic and permanent -- rural cell towers are oversubscribed and fiber will never reach mid-field sensor stations. A Gemma E2B model running on each irrigation controller node classifies soil moisture, weather forecast data cached daily, and crop growth stage to make per-zone open/close decisions autonomously, even when fully disconnected for days.
Augmentative and alternative communication (AAC) devices like Tobii Dynavox cost $8,000-15,000 and require insurance pre-authorization that takes 6-18 months, during which a nonverbal child misses the critical 2-5 year language development window — every month of delay permanently reduces their lifetime communication capacity. These commercial devices use proprietary cloud-connected software that requires annual licenses, cannot be customized by parents, and stops working when the company discontinues the model. A Raspberry Pi 5 with a touchscreen and speaker running Gemma 4 can power a fully customizable AAC device for under $75: the child taps pictograms or types partial words and the LLM predicts complete contextual sentences, learns the child's vocabulary patterns over time through local fine-tuning, and works identically at home, in the car, and in the classroom without needing WiFi. Because the model runs locally, it can be fine-tuned on the specific child's communication patterns, family names, and daily routines — personalization that a cloud API serving millions of generic users will never provide.
When a tree falls on a power line or a transformer fails, protective relays must detect the fault and trip circuit breakers within 4-16ms to prevent cascading failures that can black out entire regions -- the 2003 Northeast blackout affected 55 million people because relay coordination failed. Cloud-based AI for fault classification adds minimum 50ms network latency, which is 3-12x too slow for primary protection, meaning the fault has already propagated to adjacent circuits before the cloud model even returns a classification. The grid cannot tolerate this because electrical faults propagate at near-light speed through copper, and every millisecond of delay means more equipment damage and wider outage radius. Edge AI models embedded directly in smart relays and meters classify fault types (ground fault vs. phase fault vs. transient) in under 2ms using local sensor data, enabling intelligent protection schemes that operate entirely within the physics-mandated time window.
A 5-person machine shop running three CNC mills and a lathe generates $400K/year in revenue, but enterprise predictive maintenance platforms like Siemens MindSphere or PTC ThingWorx charge $2,000-5,000/month — nearly the profit margin of the entire shop. These platforms also require uploading proprietary machining parameters, tool paths, and production schedules to third-party cloud servers, which defense and aerospace subcontractors are contractually prohibited from doing under ITAR and CMMC regulations. An undetected spindle bearing failure at 2 AM during an unattended overnight run can destroy a $15,000 titanium workpiece and damage a $80,000 spindle. A $75 Raspberry Pi 5 with a vibration sensor running Gemma 4 fine-tuned on the shop's specific machine signatures can detect anomalous vibration patterns and halt the machine before catastrophic failure — one-time cost, no subscription, no data leaving the premises, and fine-tuned to that exact machine's normal operating profile rather than a generic cloud model trained on completely different equipment.
AI-powered tutoring apps like Khanmigo and Duolingo send every student interaction to cloud servers, consuming 50-200MB of data per hour of active use. For the 27% of US households with children who rely on metered mobile hotspots or prepaid data plans — disproportionately low-income and rural families — a child doing 30 minutes of AI-assisted homework daily burns through 3-6GB per month, which is their entire data allocation. The family faces an impossible choice: the child uses AI tutoring and the parents lose the data they need for work email and telehealth appointments, or the child goes without the educational tool that their wealthier classmates use freely on unlimited home WiFi. On-device Gemma 4 fine-tuned as a math and reading tutor operates with exactly zero data consumption after the one-time model download (2-4GB via school or library WiFi). This is not a bandwidth optimization; it is the only architecture that does not force low-income families to ration their children's education against their own connectivity needs.
Collaborative robots (cobots) working alongside humans on assembly lines must detect an approaching human hand and halt within 10ms to comply with ISO 10218 safety standards -- but cloud-based vision inference introduces 50-200ms of variable latency (jitter), meaning the robot arm cannot guarantee it will stop before striking a worker. A single collision that injures a worker shuts down the entire production line for OSHA investigation, costs $100K+ in liability, and creates lasting workforce fear of cobots that undermines adoption. Cloud APIs structurally cannot solve this because network jitter is unpredictable -- even if median latency is 30ms, the 99th percentile spike to 200ms is the one that causes the injury. An on-device vision model running on the cobot's embedded GPU delivers deterministic sub-5ms inference on every frame, and because it never touches a network, there is zero jitter variance, making safety certification possible.
In camps like Cox's Bazar (Rohingya), Kakuma (multi-ethnic East African), and Zaatari (Syrian), a single medical tent might see patients speaking Rohingya, Chittagonian, Burmese, Somali, Oromo, and Arabic in a single day, and the volunteer doctor speaks only English or French. Cloud translation APIs like Google Translate require stable internet that refugee camps in remote locations lack, and even when available, they do not support crisis-specific languages like Rohingya, which has no standardized written script. A Raspberry Pi running a fine-tuned Gemma 4 model with locally trained translation pairs for underserved languages can provide real-time spoken translation in a fully offline device costing $50 — and because the model is open-source and fine-tunable, local NGO linguists can add specialized medical vocabulary in dialects that commercial translation services will never prioritize because the market is too small to justify the engineering cost.
Manufacturing facilities running SCADA and industrial control systems enforce strict air-gap policies — no device on the factory network can reach the public internet — because a single inbound connection creates an attack vector that could shut down a production line worth $500,000/hour. When a CNC operator encounters an unfamiliar error code or alarm on a machine, they cannot pull out their phone and ask ChatGPT because the facility's Faraday-caged environment blocks cellular signals by design. They instead flip through a 400-page paper manual or wait 45 minutes for the one maintenance engineer who knows that specific machine. An on-device Gemma 4 fine-tuned on the facility's specific machine manuals, error codes, and troubleshooting procedures runs on a phone or tablet with WiFi and cellular radios disabled, fully compliant with air-gap policies. The AI must be physically incapable of making network connections to be permitted in these environments — cloud AI is not just inconvenient here, it is banned by security policy.
Wearable EEG monitors for epilepsy patients can detect pre-ictal brainwave patterns 30-90 seconds before a seizure, but only if the classification happens within 50ms of each EEG sample -- slow enough to miss the pattern and the patient falls without warning, potentially into traffic or down stairs. Cloud-based seizure detection adds 1-3 seconds of latency per inference, which collapses the detection window and produces alerts after the seizure has already begun, making them useless for fall prevention. The deeper issue is that patients wear these devices 24/7 in locations with unreliable connectivity -- subway commutes, rural areas, airplane flights -- where a cloud API call may not complete at all. A sub-1W edge chip running a quantized Gemma model processes each 50ms EEG window on-device, maintaining continuous seizure surveillance regardless of connectivity, and the data never leaves the wearable, preserving medical privacy without HIPAA-compliant cloud infrastructure.
Camera traps in wildlife reserves capture thousands of images daily, but without on-device intelligence they store everything indiscriminately — 95% of images are wind-blown vegetation or harmless animals — and rangers must physically collect SD cards from stations spread across hundreds of square kilometers on foot or by vehicle, then manually review thousands of images back at base camp. By the time a poacher image is discovered, the poacher crossed the reserve boundary 48 hours ago. Cloud-based wildlife AI services like Wildlife Insights require uploading gigabytes of images over internet connections that simply do not exist in places like the Selous Game Reserve or the Sundarbans mangrove forests. A $35 Raspberry Pi with a camera running Gemma 4 E2B can classify images on-device in real time, discard false triggers, and when it detects a human figure carrying equipment at night, immediately trigger a local LoRa radio alert to the nearest ranger station — all running on a 10W solar panel with no internet infrastructure required.
When a hurricane, earthquake, or wildfire destroys cellular infrastructure, first responders lose access to every cloud-based tool simultaneously — right when they need AI-assisted mass casualty triage, medication interaction checks, and resource allocation most urgently. After Hurricane Maria, Puerto Rico lost 95% of cell towers and had no reliable connectivity for weeks. A paramedic treating 40 injured people in a collapsed school cannot wait for Starlink to be deployed; they need decision support in the next 30 seconds. On-device Gemma 4 fine-tuned on START triage protocols, drug interaction databases, and field medicine guidelines runs entirely on the responder's phone with no infrastructure dependency whatsoever. The model must be pre-loaded and locally executable because the defining characteristic of a disaster zone is that the communication infrastructure the cloud depends on is the very thing that was destroyed.
Natural gas pipelines span thousands of miles through deserts, tundra, and mountain passes where there is no cellular or satellite internet coverage -- yet pressure and acoustic sensors along these pipelines must detect micro-leaks within seconds to prevent explosions. A 2-inch pipeline crack at 1,000 PSI escalates to a rupture in under 60 seconds, and a cloud API call that requires connectivity simply never completes in these dead zones. The structural reason cloud models fail here is not latency but total absence of network infrastructure -- no amount of API optimization helps when there is no signal. An on-device LLM fine-tuned on pipeline acoustic signatures runs on solar-powered edge nodes at each sensor station, classifying leak sounds versus environmental noise locally, and triggering automated valve shutoffs without ever needing a network connection.