Know if you will win,
before you file.
The Boston Zoning Board of Appeals has decided 15,051 development applications since 2020. PermitIQ is a machine-learning model trained on every one of them, and on the 444 hearing videos whose minutes we transcribed and parsed. For any Boston parcel, it returns a calibrated approval probability — and an auditable account of which features moved the prediction and by how much.
A typical filing costs an applicant $30,000 to $100,000 in legal, architectural, and expediter fees before the Board votes. PermitIQ is $99, used before that money is spent.
What it does.
For any Boston parcel, PermitIQ returns a calibrated probability of approval at the Zoning Board of Appeals — explained, sourced, and comparable across the city. The model is not a black box. Every prediction is accompanied by SHAP-decomposed drivers: which features moved the result, and by how much.
A snapshot of decisions from a single June 2024 hearing. Every case in the corpus has a case number, a parcel, a board, an outcome — and now, a probability we could have given the applicant in advance.
| Case | Ward | Variances | Outcome |
|---|---|---|---|
| BOA-1577058 | 22 | Height · FAR · Rear setback · Parking | APPROVED |
| BOA-1563668 | 01 | Height · FAR · Rear & side setback | APPROVED |
| BOA-1550107 | 07 | FAR · Lot area · Parking | APPROVED |
| BOA-1542615 | 21 | Height · FAR · Lot area · Setbacks | APPROVED |
| BOA-1548361 | 02 | Height · FAR · Side setback | APPROVED |
| BOA-1575505 | 12 | FAR · Lot area · Front & side setback | APPROVED |
| BOA-1554723 | 05 | Conditional use · Nonconforming | APPROVED |
| BOA-1572059 | 19 | Height · FAR · Lot area · Open space | APPROVED |
Source: City of Boston ZBA decisions, June 2024.
A worked example.
Take a four-story residential addition at 753 East Broadway, South Boston (Ward 5). The applicant intends to request a side-yard reduction and a conditional-use designation. PermitIQ's reading, computed before any filing fee has been paid:
Calibrated. The figure is 13 points below Ward 5's 86% baseline — the side-yard ask costs the applicant meaningful odds despite an otherwise favorable record at this address.
Confidence: good. Based on 412 similar cases.
| Side-yard reduction (15%) | -8 pp | |
| Ward 5 historical approval rate | +11 pp | |
| Attorney win rate, last 24 months | +7 pp | |
| Direct-abutter opposition score | -6 pp | |
| Hearing month: September | +3 pp | |
| Conditional-use designation | +4 pp |
Methodology.
The model is a gradient-boosted classifier (XGBoost) trained only on signals known before a hearing — the variances requested, use and project type, whether an attorney is engaged, and ward / zoning historical approval rates. It is isotonic-calibrated so the probabilities are meaningful, and validated with parcel-grouped cross-validation — no parcel appears in both training and test, so the score reflects genuinely unseen filings. Forward AUC: 0.62 (held-out 0.70). Brier score: 0.06. With a ~90% city-wide approval base rate this is a directional risk signal — it reliably separates higher- from lower-risk filings and shows you why, but it is not a guarantee.1
- Boston ZBA decision letters (OCR via Google Cloud Vision)
- Hearing video transcripts (Whisper)
- Boston Open Data: parcels, permits, violations, complaints, crime, transit, census
- FEMA flood zone & sea-level-rise overlays
- Property assessment records (FY 2026)
- Cherry-Hill ZBA tracker (filing dates, hearing dates)
The feature set spans ward history, attorney track record, variance complexity, zoning context, neighborhood opposition, and transcript-derived hearing-climate signals. Sixteen candidate features were dropped at training time after they were identified as post-hearing leakage — they encoded information the applicant could not have known before the Board voted.2
Every prediction surfaces the top eight SHAP-decomposed drivers with direction and magnitude. The applicant sees not only the probability, but a defensible account of why the model arrived there. The expectation is informed disagreement, not blind trust.
The Boston Open Data feeds and the ZBA decision corpus refresh every Sunday at 2 AM ET. The model retrains automatically when new decisions arrive, and the calibration is checked on rolling windows.
1 AUC = area under the receiver-operating-characteristic curve. Brier = mean squared error of predicted probabilities against outcomes. ECE = expected calibration error.
2 The leakage analysis is documented in the project README; the dropped features are listed by name in train_model_v2.py.
How to use it.
Three modes of use, in roughly increasing depth:
- iA single parcel prediction.
Type an address. Receive a calibrated probability with the driving SHAP features. Free, no signup. /predict →
- iiA full hearing-prep report.
Adds opposition risk for the neighborhood, attorney recommendations ranked against your specific variance types, board member voting tendencies for the panel hearing your case, and filing-timing analysis. $99 per parcel. /intelligence →
- iiiProgrammatic access for portfolios.
For developers, brokers, and law firms working across many parcels. REST API, batch endpoints, white-label reports, audit logs. Pro from $499/mo; Agency by conversation. /pricing →
The Boston Zoning Digest.
One short letter, every Friday.
New ZBA decisions, attorney win-rate movement, parcels whose approval odds shifted this week, and the variances most likely to be denied next cycle.
One report could save you fifty thousand dollars.
A typical Boston ZBA filing costs $30K – $100K in legal, architectural, and expediter fees. PermitIQ is $99, used before any of that is spent.
PermitIQ provides statistical risk assessment of public zoning data. It is not legal advice, does not guarantee any outcome, and is not a substitute for engaging a qualified Boston zoning attorney. The model reports calibrated probabilities; individual cases vary.