How our AI works

Seven engines, each shown as inputs → steps → outputs, with the checks, limits and measures around it. Each links to a working demo where one exists.

1 · Bill reader and system sizing

Turns a photo of an electricity bill into a sized system, subsidy estimate and payback. Run it →

InputBill photo or PDF, city, roof area, sanctioned load
1 · OCRImage cleaned, text recognised
2 · ExtractPattern and layout rules find units, amount, tariff, consumer number
3 · ValidateCheck amount ≈ units × tariff; flag outliers
4 · SizePhysics model: demand vs roof vs load limits
5 · MoneySubsidy, savings, payback, EMI

The sizing model is plain physics and arithmetic, not a guess:

kW needed   = monthly units ÷ (kWh produced per kW per month in that city)
kW by roof  = shade-free sq ft ÷ ~100
system kW   = min(needed, roof, sanctioned load, 10 kW), 0.5 kW steps
generation  = kW × yield × 12 × (1 − 0.005)^(year−1)
savings     = min(generation, consumption) × tariff × (1.03)^(year−1)

Subsidy follows the publicly announced PM Surya Ghar structure (₹30,000/kW for the first 2 kW, ₹18,000 for the third kW, capped at ₹78,000) and is checked against the current scheme before any quote is sent.

Where AI is usedOCR and field extraction from messy images; later, roof detection from satellite or drone imagery.
Human checkExtracted fields are shown for confirmation. Low-confidence reads go to a person.
Known limitsBills differ by DISCOM and language; blurry photos fail. Yield is a city average, not a site simulation. Excludes inverter replacement.
Metrics we trackField extraction accuracy per DISCOM; proposal-to-order rate; gap between quoted and actual generation.

2 · Lead scoring

Predicts how likely a homeowner is to buy so that follow-up goes to the best leads first. Run it →

InputBill size, ownership, roof, shade, timeline, area, bill uploaded
1 · FeaturesEach answer becomes a number
2 · WeightsFactors add or subtract points
3 · SigmoidTotal mapped to 0–100%
4 · ExplainPer-factor contribution shown
OutputScore, tier, reasons
score = −0.5 + Σ weight(factor)          (demo: hand-set weights)
probability = 1 ÷ (1 + e^(−score))

This is a logistic model, the simplest explainable classifier. Because every contribution is visible, a salesperson can see why a lead scored low and fix it, for example by asking for the bill.

Scoring is separate from billing. A lead is billable only when fixed contract conditions are met (bill verified, in area, roof sufficient, visit booked, owner consent). A high score never creates a charge by itself.

Production pathRetrain the same structure on real outcomes (sold / not sold) once we have a few hundred closed leads per city. Check calibration so "70%" really means about 7 in 10.
Human checkScores only reorder follow-up. No lead is discarded on score alone.
Known limitsWeights in the demo are illustrative. Seasonality and local events change behaviour. Small samples give noisy models.
Metrics we trackAUC and calibration by city; lead-to-visit and visit-to-order rates by tier.

3 · Document checker

Finds the usual causes of subsidy and net-metering rejection before you file. Run it →

InputApplication fields and uploaded documents
1 · ReadOCR of bill and ID text
2 · Cross-checkNames, numbers, dates across documents
3 · RulesDISCOM- and portal-specific checks
OutputReadiness score and fix list

Fuzzy name matching compares names after removing titles (Mr, Shri, Smt) and punctuation, sorting words, and measuring edit distance. It also treats initials as a match, so "R K Sharma" and "Rakesh Kumar Sharma" are flagged as probably the same person, confirm spelling rather than failing outright.

similarity = 1 − editDistance(sorted(nameA), sorted(nameB)) ÷ maxLength
≥ 92% pass · 60–92% warn · < 60% fail

Rule checks include mobile format, pincode, consumer number present, bill age, tenant needs owner NOC, proposed capacity versus sanctioned load, and the required document list. Rules are configured per DISCOM because requirements differ and change.

Human checkThe checker advises; a person submits. It never files or rejects an application by itself.
Known limitsRules can go stale when portals change. OCR can misread poor scans. We cannot see the official system's own validation.
Metrics we trackFirst-time approval rate with and without the checker; average days to approval; top rejection reasons by DISCOM.
Data careIDs and addresses are sensitive. See Privacy: minimised, encrypted, and deleted on schedule.

4 · WhatsApp qualification agent

Holds a short conversation in English or Hindi, collects what an installer needs, and books a visit. Chat with it →

InputHomeowner message or button tap
1 · LanguageDetect script and language
2 · IntentQuestion (subsidy, cost, loan) or answer?
3 · SlotsExtract city, ownership, bill, roof, timeline
4 · DecideNext question, FAQ, or handoff
OutputReply plus structured lead record

The agent is a slot-filling dialogue system. It keeps a record of what it still needs, asks for the next missing item, and can be interrupted by questions without losing its place. In the demo, intent detection is keyword-based so you can read every decision in the "what the agent is thinking" panel.

Production path: a large language model handles free-form wording and local-language variation, but is kept inside guardrails: it may only answer from an approved knowledge base (scheme rules, FAQs), must never invent prices or promise subsidy, and fills the same structured slots so the rest of the system is unchanged.

GuardrailsApproved answers for money topics; estimates always labelled as estimates; opt-out honoured on request ("STOP").
Human handoffAsked for a person, angry, confused twice in a row, or high-value lead: a human takes over with the full transcript.
Known limitsCode-mixed Hinglish, voice notes and unusual phrasing need an LLM and testing per language. Meta's WhatsApp Business rules apply to outbound messages.
Metrics we trackCompletion rate; handoff rate; wrong-slot rate (audited sample); lead-to-visit rate; opt-out rate.

5 · Pay-on-proof ledger

The mechanism that makes "pay only for results" auditable. Run it →

EventLead qualifies, file approved, order signed…
1 · EvidenceSystem attaches the proof (bill, portal status, signed PDF)
2 · RuleContract definition tested automatically
3 · EntryTimestamped ledger line and fee
4 · DisputeEither side can flag within the window
OutputMonthly invoice with evidence

This engine is deliberately not AI. Billing decisions are deterministic rules applied to recorded evidence, so every charge can be explained and reproduced. AI proposes (this lead looks qualified); the rule decides (all five conditions are met, with proof).

Why it mattersInstallers can verify every line, which is the basis of trust in a result-based model.
DisputesDisputed lines are paused from the invoice and reviewed by a person. See Payments & Disputes.
Known limitsDepends on installers recording outcomes honestly. Non-circumvention terms and periodic audits cover that.
Metrics we trackDispute rate; time to resolve; share of disputes upheld.

6 · Performance monitoring (roadmap)

For installed systems: find lost generation and turn it into service revenue. Not in the demo yet.

InputInverter or app data, location, tilt, capacity
1 · ExpectedWeather and irradiance model gives expected kWh
2 · ComparePerformance ratio = actual ÷ expected
3 · DetectAnomaly detection on the gap pattern
4 · DiagnoseSoiling, shading, fault, outage
OutputAlert and recommended action
performance ratio (PR) = actual kWh ÷ expected kWh (weather-corrected)
gradual PR decline after rain-free weeks → soiling → schedule cleaning
sudden step drop → string or inverter fault → dispatch technician
daily dip at the same hour → shading from a new obstruction

Recovered generation is measured against an agreed baseline, which is what the AMC fee in pricing is based on.

Known limitsInverter brands use different data formats; many homes have no monitoring at all. Baselines must be agreed in advance.
Metrics we trackFalse-alert rate; verified recovered kWh; cleaning visits that restored output.

7 · Safeguards around all of it

Human in the loop

AI never rejects an application, never decides billing alone, and hands off to a person when unsure.

No automated promises

Subsidy, savings and payback are labelled estimates. Final figures come from the installer's survey and official approvals.

Data minimisation

We collect what a task needs, not more. Identity documents are used for the application only. See Privacy Policy.

Fairness checks

Scoring never uses religion, caste, gender, or similar attributes. We review outcomes by area and language for unintended skew, such as under-serving Hindi speakers or smaller towns.

Auditability

Every model version, rule set and decision input is logged, so a score or alert can be explained later.

Measured, then claimed

We test on real outcomes before stating any accuracy. Until then this site makes no accuracy claims.