Community health workers in remote clinics across the Amazon, rural India, and sub-Saharan Africa serve as the only medical decision-makers for populations of 2,000-5,000 people, but they have limited training and no access to specialist consultations or cloud-based diagnostic tools because satellite internet costs $100+/month — more than their entire monthly salary. When a child presents with a high fever, the health worker must decide whether it is malaria, dengue, or a simple infection, and getting it wrong means either a 6-hour canoe ride to the nearest hospital for a false alarm or a preventable death from delayed treatment. A Raspberry Pi running Gemma 4 fine-tuned on WHO clinical decision protocols can walk the health worker through a structured triage questionnaire in their local language, suggest probable diagnoses, and flag red-flag symptoms that require emergency evacuation — all for a one-time $50 hardware cost with no recurring fees and no data leaving the device, which matters because patient health data in these communities is highly sensitive and covered by emerging data sovereignty laws.
Real problems worth solving
Browse frustrations, pains, and gaps that founders could tackle.
CNC mill spindle bearings exhibit vibration signature changes 100-300ms before catastrophic failure, but cloud-based anomaly detection requires 400-600ms round-trip (sensor to cloud API to response), meaning the alert arrives after the $50,000 spindle is already destroyed. A single undetected spindle crash halts the production line for 8-12 hours while the spindle is replaced, costing a mid-size shop $15,000-$40,000 in lost production on top of the part cost. Cloud API models cannot solve this because the physics of network latency are immutable -- even with 5G, the serialization, transmission, inference, and return path exceeds the failure window. An on-device model like Gemma 4 E2B running on a vibration sensor gateway processes accelerometer data in under 5ms and triggers an emergency spindle stop before the bearing seizes, because the inference happens on the same board that reads the sensor.
Migrant agricultural workers — many earning below minimum wage — receive pay stubs, employment contracts, and safety documents in English they cannot fully read, and are routinely underpaid or exposed to hazardous conditions because they cannot verify what they signed. Cloud-based translation and document analysis APIs cost $20+/month per user, which is prohibitive when you earn $12/hour picking strawberries for 6 months a year. Even free tiers require account creation with email verification, credit cards, and terms-of-service agreements — all barriers for workers with limited digital literacy and no stable address. An on-device Gemma 4 model with multilingual capability (trained on 140+ languages) loaded once onto a cheap Android phone can translate and explain employment documents indefinitely with zero ongoing cost, zero account requirements, and zero data collection. The structural requirement is that the AI must be permanently free to operate after initial setup, which is architecturally impossible with API-based models that meter every token.
700 million smallholder farmers worldwide grow crops in areas with no reliable internet, yet crop disease identification requires expert knowledge that most farmers lack — a single misdiagnosis (e.g., confusing bacterial wilt with fusarium) means applying the wrong treatment, which wastes scarce pesticide money and lets the real disease destroy the entire season's yield. Cloud-based AI plant diagnosis apps like Plantix exist but are useless when the farmer's phone shows 'No Service' for weeks at a time, and even when connectivity briefly appears, uploading a photo over 2G takes minutes and costs a meaningful fraction of daily income in data fees. A $35 Raspberry Pi Zero with a camera module running a fine-tuned Gemma 4 E2B model can identify 50+ crop diseases from leaf photos in under 2 seconds with zero connectivity, and because the model is fine-tunable, local agricultural extension offices can retrain it on region-specific diseases and local crop varieties that global cloud models have never seen in their training data.
Cloud-based visual assistance apps like Be My Eyes and Seeing AI require constant connectivity to describe surroundings to visually impaired users, but the places where blind users most desperately need real-time scene description — underground subway stations, elevators, parking garages, building stairwells — are exactly the places with zero cell signal. A blind person navigating a New York subway transfer between the A and C trains at 59th Street has no connectivity for the 4-7 minutes they are underground, precisely when they need an AI to read platform signs, identify which train is arriving, and describe obstacles. Cloud AI fails at the exact moment the user's safety depends on it. On-device Gemma 4 with multimodal vision capability can continuously process the phone's camera feed and provide audio descriptions entirely offline — no connectivity gaps, no latency spikes, no service interruptions. The model must be on-device because the environments where blind users face the highest navigation risk are structurally the same environments where internet is unavailable.
Delayed Auditory Feedback (DAF) therapy for stuttering requires playing back the speaker's own voice with a precisely controlled 50-200ms delay — this is the therapeutic mechanism that helps the brain reorganize speech timing. Cloud-based speech processing adds 200-500ms of network round-trip latency on top of the therapeutic delay, pushing total feedback delay past 400ms where the technique stops working and actually worsens disfluency. Every millisecond of uncontrolled jitter in cloud latency corrupts the therapeutic signal. On-device inference with Gemma 4's native audio processing can maintain sub-20ms processing latency, giving the app precise control over the total delay window. The physics of speech therapy demand that the AI processing happens on the same device as the microphone and speaker — any network hop between them breaks the therapeutic mechanism. This is why no cloud-based stuttering app has matched the efficacy of dedicated $3,000 DAF hardware devices.
When a journalist in Russia, Iran, or China uses a cloud AI service to help analyze leaked documents, translate whistleblower communications, or research a sensitive story, the API call traverses state-controlled internet infrastructure where deep packet inspection can identify the query destination, and the cloud provider may be legally compelled to hand over logs to local authorities. Even using a VPN, the journalist creates a pattern of encrypted traffic to known AI endpoints that itself becomes suspicious metadata. The consequence is not abstract: Russia has used AI-based facial recognition to arrest journalists, and source identification from digital traces has led to imprisonment and worse. An on-device model running entirely offline on a phone in airplane mode produces zero network traffic, zero server-side logs, and zero metadata — the journalist can analyze documents, draft stories, and translate sources without creating any digital evidence that the work occurred. This is not a privacy preference; it is a physical safety requirement.
Insurers deny or retroactively recoup claim payments 90-180 days after a service was rendered and initially approved, citing post-payment audit findings like coding technicalities or eligibility verification gaps that the provider had no way to catch at time of service. A small pediatric practice delivers a round of vaccinations, receives payment, then three months later gets a demand letter clawing back $8,000 because the insurer retroactively determined the patient's coverage had a 2-day gap. The practice has already paid the nurse, bought the vaccines, and reported the revenue. This persists because insurer-provider contracts typically grant insurers 12-24 month clawback windows with no reciprocal right for providers to rebill after the same period, and small practices lack the legal resources to challenge recoupment demands.
A solo-practice therapist who wants to use an LLM to help structure session notes, generate treatment plans, or draft insurance pre-authorization letters cannot send patient details through OpenAI or Anthropic APIs without a Business Associate Agreement — and even with a BAA, every API call creates a copy of PHI on a third-party server that becomes a breach liability. If a therapist pastes 'Patient reports suicidal ideation following divorce from [name]' into ChatGPT, that data may persist in logs indefinitely, and a single breach can trigger $50,000-per-violation HIPAA penalties that would bankrupt a solo practice. On-device Gemma 4 fine-tuned on clinical documentation templates processes the note entirely on the therapist's phone — PHI never leaves the device, no BAA is needed, no server logs exist, and the therapist gets AI-assisted documentation without creating a regulatory time bomb. The structural requirement is that the model runs where the data already lives: on the clinician's own hardware.
Despite the Mental Health Parity and Addiction Equity Act requiring equal coverage for mental and physical health, insurers routinely impose stricter utilization review on therapy visits than on comparable medical visits, effectively capping therapy at 20-30 sessions per year by requiring re-authorization every 6-8 visits with increasingly burdensome clinical documentation. A patient in treatment for PTSD finds their therapist spending 45 minutes per session on paperwork justifying continued treatment, time that could be spent on actual therapy. The DOL found parity violations in 73% of audited plans, yet enforcement actions are rare. This persists because the law requires "comparable" management of mental and medical benefits but does not define quantitative thresholds, giving insurers room to argue their review processes are technically equivalent while in practice being far more restrictive for behavioral health.
Over 2.6 billion people globally lack internet access, and community health workers (CHWs) serving remote villages in sub-Saharan Africa, rural India, or Appalachia cannot use cloud-based AI triage tools when they are actually standing in front of a sick patient. A CHW in rural Malawi who suspects pneumonia in a child needs decision support right now — not when they walk 8 kilometers back to a town with cell coverage. Cloud-based medical AI is structurally useless here because the point of care and the point of connectivity never overlap. An on-device model fine-tuned on WHO IMCI protocols and loaded onto a $150 Android phone can provide symptom-based triage guidance with zero connectivity, running inference locally in under 2 seconds. This is the only architecture that puts AI capability at the actual moment of clinical decision-making in low-resource settings.
Insurer online provider directories list physicians as in-network and accepting patients when they have actually left the network, moved practices, or retired, leading patients to book appointments and receive care under the false assumption of in-network coverage. The patient only learns the provider was out-of-network when they receive a bill 4-8 weeks later, sometimes for thousands of dollars, and their only recourse is a lengthy appeals process with a low success rate. A CMS secret shopper study found that 52% of provider directory listings contained at least one inaccuracy. This persists because insurers face no financial penalty for directory errors under most state laws, directory updates depend on providers self-reporting changes, and there is no shared national credentialing database that syncs in real time.
When a DV survivor asks ChatGPT or Claude for help drafting a safety plan, finding shelters, or understanding custody law, those queries are logged on company servers and can be subpoenaed in custody battles or divorce proceedings. An abuser's attorney can compel OpenAI or Anthropic to produce chat logs showing the survivor was planning to leave, which gets reframed as 'premeditation' or 'parental alienation' in court. The survivor is forced to choose between getting no AI help at all or creating a discoverable paper trail that their abuser's lawyer will weaponize. On-device models like Gemma 4 running locally on a phone produce zero server logs, require no account creation, and leave no network trace — the conversation exists only in local memory that the survivor can wipe instantly. This is not a 'nice to have' privacy feature; it is the difference between a survivor safely planning an exit and an abuser's attorney presenting exhibit A.
Insurers increasingly use "copay accumulator adjustment" programs that accept manufacturer copay assistance cards at the pharmacy but refuse to count those payments toward the patient's deductible or out-of-pocket maximum. A patient on a specialty drug for multiple sclerosis uses a $15,000/year copay card, assumes they are building toward their $4,000 deductible, then in month 5 when the copay card exhausts, discovers they still owe the full $4,000 deductible and face paying full price for the remaining months. This effectively doubles the patient's annual cost while the insurer pockets both the manufacturer's copay assistance and the patient's deductible. It persists because there is no federal regulation requiring transparency about accumulator programs, and insurers bury the policy in plan documents that virtually no enrollee reads before selecting a plan.
Insurers require "step therapy" (fail-first) protocols where a patient must try and fail on one or two cheaper medications before the insurer will cover the drug their doctor actually prescribed, even when the doctor has clinical reasons to skip the cheaper options. A rheumatoid arthritis patient whose doctor prescribes a biologic must first spend 3-6 months on methotrexate, experiencing documented liver toxicity risks, before the insurer approves the biologic. This costs the healthcare system more in the long run because failed therapies generate additional office visits, lab tests, and ER visits. Step therapy persists because PBMs negotiate rebates on preferred drugs, and every month a patient stays on the cheaper drug saves the insurer money regardless of patient outcomes, and override requests require physicians to spend 30+ minutes on peer-to-peer calls.
Insurers in ACA marketplace plans list specialists in their provider directory who are not accepting new patients, have retired, or are located 90+ miles away, technically satisfying network adequacy requirements while leaving enrollees with no practical access to care. A patient in rural Texas who buys a Silver plan expecting dermatology coverage discovers the only listed dermatologist is 3 hours away and booked 8 months out, effectively making the coverage useless. State regulators require insurers to submit provider directories annually, but directories are stale within weeks because there is no real-time verification system and no penalty for inaccuracy. Insurers benefit from "ghost networks" because narrow networks reduce claim volume, and enforcement requires individual patient complaints that most people never file.
When a property suffers damage (burst pipe, fire, break-in), the insurance adjuster requires pre-loss documentation: appliance serial numbers, receipts for recent renovations, photos of the unit's condition, and proof of maintenance history. Most managers have none of this organized — appliance info is on a sticky note in the unit file, renovation receipts are in the owner's email, and the last inspection photos are on a former employee's phone. Without pre-loss documentation, claims are reduced or denied, leaving owners to absorb $5,000-$50,000+ in repair costs. This persists because there is no standard onboarding checklist that links asset documentation to insurance requirements, and managers do not get rewarded for preventive documentation — they only feel the pain when a claim is filed.
After every medical visit, insurers mail an Explanation of Benefits (EOB) that uses proprietary billing codes, nested adjustment columns, and ambiguous "you may owe" language that even physicians and billing specialists cannot parse reliably. Patients cannot tell from the EOB whether they owe money, how much, or whether the claim was processed correctly, so they either ignore legitimate bills (which go to collections) or pay inflated amounts without questioning. A 2023 JAMA study found that 80% of medical bills contain errors, but patients lack the literacy to catch them. This persists because EOB formatting is not standardized by CMS for commercial plans, and insurers have no incentive to make them clearer since confused patients are less likely to appeal underpayments or catch billing mistakes.
In multifamily properties without individual meters, managers must allocate shared utility costs (water, gas, trash) to tenants using RUBS (Ratio Utility Billing System) based on unit square footage, occupancy, or some hybrid formula. Tenants receive a monthly charge with no breakdown of the master meter reading, the allocation formula, or how their share was computed. The result is that 10-20% of tenants dispute their utility bill-back every single month, each dispute consuming 15-30 minutes of manager time. This persists because RUBS calculations are done in Excel with no tenant-facing transparency layer, and utility billing vendors charge $3-$8 per unit per month — a cost owners resist when already paying the manager 8-10% of gross rent.
Property managers of condo units in HOA-governed communities receive violation notices (unapproved paint color, trash bin visible, holiday decorations left up) weeks or months after the alleged infraction, often without photographic proof or a specific date. The manager must then confront the tenant about something neither party remembers, the tenant denies it, and the fine escalates. Owners pay $50-$500 in fines for violations they cannot verify, and managers waste hours mediating disputes between tenants and HOA boards. This persists because HOA boards rely on volunteer drive-by inspections with inconsistent documentation standards, and there is no shared platform between HOA management companies and property managers for real-time violation tracking.
Health insurers reclassify drugs between formulary tiers mid-plan-year, moving a patient's stable medication from Tier 2 ($30 copay) to Tier 4 ($150 copay) or off-formulary entirely with no clinical justification. A patient who has been stable on a specific antidepressant for two years is suddenly forced to either pay 5x more or switch to a "preferred" alternative that may cause different side effects and require weeks of dose titration. Insurers are allowed to do this because CMS only locks formularies for Medicare Part D mid-year; commercial plans face no such restriction. Pharmacy benefit managers (PBMs) drive these changes to capture rebate revenue from competing drug manufacturers, meaning the tier assignment reflects financial deals, not clinical efficacy.
Every time a tenant moves out, property managers must re-key or replace locks on all entry points to ensure the prior tenant's copies cannot be used. For a 200-unit portfolio with 40% annual turnover, that is 80 re-keying jobs at $150-$300 each, totaling $12,000-$24,000 per year — plus the coordination overhead of scheduling a locksmith during the narrow turnover window. If re-keying is delayed even one day past move-in, the new tenant has a legitimate safety concern and the manager faces liability. This persists because traditional pin-tumbler locks require physical re-keying by a licensed locksmith, and smart lock adoption in rental housing is slow due to upfront hardware cost ($200-$400/door), tenant tech-literacy concerns, and lack of integration with existing property management software.
When a patient schedules surgery at an in-network hospital, the anesthesiologist, pathologist, or radiologist assigned to their case is frequently out-of-network, resulting in a surprise bill averaging $1,219 that the patient only discovers weeks after the procedure. Patients cannot choose or even know who these providers will be beforehand, so there is no way to shop around or verify network status. The No Surprises Act (2022) addressed emergency situations but left gaps for scheduled procedures in certain states, and enforcement is fragmented across state insurance commissioners with tiny audit staffs. This persists because hospitals contract with independent physician staffing groups that deliberately stay out of major networks to maximize reimbursement rates, and insurers have no leverage to force them in-network since the hospital controls facility access.
After receiving an application, property managers must verify income (request pay stubs, call employers), run credit and background checks (often through separate services), and contact 1-2 prior landlords who frequently do not answer the phone or respond to emails for days. During this 5-7 day screening window, qualified applicants accept other units, leaving the manager to restart with weaker candidates. In competitive rental markets, this slow screening directly causes longer vacancies. This persists because prior landlord verification has no standardized protocol — there is no registry of property managers, no API to query rental history, and many prior landlords are themselves small operators with no obligation or incentive to respond promptly.
Most health insurers still require prior authorization requests for advanced imaging (MRIs, CT scans) to be submitted via fax, which then cycle through 2-3 internal review queues before a determination is made. This means a patient with suspected torn ACL waits two weeks in pain before they can even get the scan that confirms whether they need surgery, turning a 4-week recovery timeline into 6+ weeks. The delay exists because insurers built their utilization management systems in the 1990s around fax-based workflows, and re-platforming would require retraining thousands of nurse reviewers and rewriting contracts with delegated review vendors. The AMA reports that 94% of physicians experience care delays due to prior auth, but no single insurer has incentive to modernize unilaterally because slow approvals reduce short-term claim payouts.
An estimated 60-70% of wedding venues require couples to select catering, bar service, and sometimes florals and lighting exclusively from a "preferred vendor" list of 3-5 companies that pay the venue referral fees of 10-20% of contract value. This eliminates price competition and inflates per-plate costs by $15-$40 compared to open-market catering. For a 150-person wedding, that's $2,250-$6,000 in hidden venue-vendor kickback costs passed directly to the couple. Couples cannot negotiate because the mandate is buried in the venue contract they already signed, and switching venues means losing their deposit and their date. The system persists because venues earn significant passive revenue from vendor kickbacks — often $5,000-$15,000 per wedding — creating a closed marketplace where vendors compete on referral fee size rather than price or quality.
Most property managers begin lease renewal outreach 60 days before expiration, but tenants often need 2-3 rounds of negotiation on rent increases, and if they decline, the manager needs 45+ days to market, show, screen, and move in a new tenant. Starting too late means a unit sits vacant for 30 days at $2,000/month rent — a $2,000 loss that dwarfs any rent increase the manager was negotiating. Multiply this across a 100-unit portfolio and late renewal outreach costs $20,000-$60,000 annually in preventable vacancy. This persists because managers track lease dates in spreadsheets or calendar reminders that fire once, with no automated escalation sequence or decision tree for different tenant response scenarios.
Wedding insurance costs $200-$600 and covers weather, vendor no-shows, and property damage — but virtually every policy explicitly excludes "change of heart" or relationship dissolution, which is the single most common reason weddings get canceled outside of pandemics. Couples buying insurance assume they're protected against cancellation, discover the exclusion only when filing a claim, and lose $20,000-$50,000 in non-refundable vendor deposits with no coverage. This matters because it's insurance that doesn't cover the primary risk, which is like buying fire insurance that excludes kitchen fires. The exclusion persists because insurers cannot underwrite relationship stability and adverse selection would be extreme — couples most likely to cancel are most likely to buy — so the product is marketed on a comforting feeling rather than actuarial reality.
When a property manager dispatches a plumber or HVAC tech for an urgent repair, the vendor frequently no-shows or reschedules last-minute because small residential jobs are low-priority compared to commercial contracts. The manager only finds out when the tenant calls to complain, then scrambles to find a backup vendor at emergency rates. For habitability issues like no heat in winter, this can trigger legal liability under implied warranty of habitability statutes, exposing the owner to rent abatement claims or code enforcement fines. This persists because property managers maintain a mental rolodex of 3-5 vendors per trade with no real-time availability data, and vendors have no penalty for no-showing on residential work since there is always more demand than supply for licensed tradespeople.
Being asked to be a bridesmaid in the U.S. now carries an average out-of-pocket cost of $1,200-$2,500 per person: a specific dress ($150-$300), shoes ($80-$150), hair and makeup ($150-$300), bachelorette trip ($300-$1,000), shower gift, wedding gift, and various other line items the bride selects unilaterally. Declining any individual expense risks social fallout and "ruining the aesthetic." This is painful because these costs disproportionately hit young women in their mid-20s who are least able to absorb them, and saying no to the honor means losing a friendship. The dynamic persists because wedding culture has escalated "bridesmaid duties" into a performative spending competition amplified by social media, and there is no socially acceptable script for saying "I can't afford this" without it being read as "I don't care enough."