What changed
FACT: EPA finalized the Lead and Copper Rule Improvements (LCRI), requiring drinking-water systems to inventory and replace lead and certain galvanized service lines and lowering the lead action level to 0.010 mg/L (federalregister.gov 2024-23549). FACT: frontier multimodal models became cheap and fast enough to read messy handwritten/scanned records at scale (GPT-5.6 economics — 2.2x faster, 27% cheaper). HYPOTHESIS: the two together make afternoon-scale digitization of decades of paper tap cards economically viable for the first time.
Why now
A finalized federal mandate with hard compliance obligations meets a step-change in cheap vision inference. Every affected system must produce and maintain a service-line inventory in a defined schema; thousands of small/rural systems have paper records and no data staff to key them by hand.
Converging signals
Regulation (LCRI forced inventory/replacement) × AI capability (cheap handwriting/scan OCR). The rule, the defined filer class (community water systems), and the state submission portal are three signals meeting at one point — that is the convergence.
Customer pain
HYPOTHESIS (not in source, must validate): small utilities face months of manual key-entry from tap cards and permit records to build an LCRI inventory, and lack in-house staff to do it; engineering firms doing it bill hourly. FACT: the obligation and material-classification requirement are real per the LCRI.
Who pays
Small and rural community water systems that must inventory/replace lead & galvanized lines, and — more reachably — the engineering/consulting firms already contracted to do these inventories for dozens of systems each.
Solved today
HYPOTHESIS: hand-keying by staff or interns, spreadsheets, or an engineering firm billing hourly to transcribe records; some GIS/asset-management vendors bolt on inventory modules. Needs verification.
Why current solutions are bad
Manual transcription is slow, error-prone, and expensive per connection; it does not scale to the number of small systems facing the same deadline, and staff-less utilities cannot do it at all.
Proposed product
Upload/scan tap cards + permits → vision model transcribes and normalizes → rules layer applies LCRI material-classification logic → web app for human review/correction → export in the state's required inventory format. Priced per connection/record.
MVP version
Ingest a batch of scanned tap cards for ONE system, transcribe with a vision LLM, apply a material-classification ruleset, produce a reviewable table + one state's export. Human-in-the-loop correction UI is mandatory — the output feeds a regulatory filing.
30-day build
Obtain 500 real scanned tap cards from one friendly utility/engineering firm; build the transcription+classification pipeline; measure accuracy against an engineer-of-record's ground truth (the kill test). Encode one state's inventory schema.
60-day build
Add review/correction UI and export; run a paid pilot with one engineering firm across 2-3 systems; document accuracy and time saved vs manual.
90-day revenue plan
Sell per-connection digitization to the pilot firm and 2-4 more small systems; publish an accuracy/time-savings case study; add a second and third state's export format.
Distribution path
Direct outreach to engineering/environmental consulting firms doing LCRI inventories (they aggregate many systems = one sale, many connections), state rural water associations, AWWA sections, and state drinking-water program lists. Value-demonstration (case study) not relationship sales.
Pricing hypothesis
Per-connection/record fee (e.g. a few dollars per line digitized) with a per-system minimum; or a flat per-system inventory package. Firm-level volume tier.
Technical difficulty
Moderate. OCR of messy handwriting is the risk; the LCRI rules layer and export formats are tractable. Human review contains the accuracy risk.
Legal / regulatory risk
Low-moderate. Output feeds a regulatory filing, so accuracy/liability must be managed — position as a preparation tool reviewed and signed off by the engineer/utility of record, not the certifier of record. No licensure required to transcribe records.
Platform dependency
Low. Submits to government inventory systems — no platform owner can deplatform it. Dependence on a commercial LLM API is a cost/availability consideration, not a policy risk.
Founder fit
Very high. This is the founder's proven shape: a federal mandate forces a defined class to file with a government system, and a solo operator builds the submission/preparation layer and charges per transaction — directly analogous to his shipped FMCSA ELDT registry app. Public-records + industrial-ops strengths apply.
Breakout potential
High and durable: the rule is nationwide and permanent, ~50 state formats to replicate into, and ongoing inventory-maintenance/annual-update obligations create recurring rather than one-time revenue.
Final recommendation
PURSUE, gated on the kill test. Highest-founder-fit shape (mandate → forced filer → per-transaction tool), durable nationwide + 50-state demand, reachable buyer via consulting firms. Before any real build, run the 500-real-card accuracy test with an engineer of record and confirm remaining greenfield demand vs. already-completed inventories.
Next action
Secure ~500 real scanned tap cards from one utility or LCRI-serving engineering firm and benchmark transcription + material-classification accuracy against their verified ground truth.