LeadPolish cleans a CSV or Excel export, merges the duplicates, verifies every email address and phone number, then grades each record on whether the person behind it is actually worth someone's afternoon. It runs in memory on your own machine. There is no database, no account, and nothing to delete afterwards.
Every number below comes from benchmarks/run.py against 200
hand-labelled leads. The labels were written from what a person would say
after ten seconds on each row, before any of the scoring existed.
200 synthetic leads built to look like a real export: Indian and international
names, phone numbers in six formats, duplicates with the surname misspelt,
shared info@ mailboxes, disposable addresses, placeholder junk,
and free-mail addresses filed under company names. Synthetic data cannot
prove behaviour on your list — it can only show the tool does what it
claims on data whose right answer is known. Run it yourself:
python benchmarks/run.py.
Bad encodings, semicolon delimiters, BOMs, blank spacer rows, ragged lines,
Excel turning a phone number into 9.82e+09. Columns are detected
from the headers and the cell contents, so a column called
Column3 full of things with an @ in it is still the
email column. Anything it does not recognise passes through untouched.
Every proposed change is shown as before → after. Every duplicate group carries its score and its reasoning, and you approve or reject them one at a time. Two records with a different email and a different phone number are never merged on a similar name — they are surfaced for you to judge instead.
Each record is scored on four things: can you reach them, can they say yes, is this a real business contact, and do they match the buyer you described. Out comes an A–D grade, a next step, and the reasoning in plain English. Describe your buyer in one click and the ranking rearranges to match.
A cleaned CSV or Excel file with the grades written in, plus a PDF report that opens with a plain-English paragraph, shows real before/after evidence rather than only counts, and pairs every issue with what to do about it. Put your own name and logo on it.
Every tool checks syntax and MX records. LeadPolish also compares the email
domain against the website and the company name. So
minh.nguyen@vnexpress.net filed under "Minh Corp" gets flagged:
the address works, it just belongs to a newspaper. That record passes every
other validator on the market.
A generic lead score answers a question nobody asked. Say you sell to marketing directors in India at registered companies, and the same file reorders itself around that. Records that break a rule are still scored and ranked, and marked so you can filter them out — never silently dropped.
The engine is deterministic: same file, same result, no key, no network calls beyond an optional DNS lookup. Paste in your own provider key and the AI gets only what the rules deliberately refused to guess at — the borderline duplicate pairs and a column's distinct values. If the key is wrong or the provider is down, those steps are skipped and the rest is byte-identical.
No database, no session, no cache, no temp files. The upload is held in memory for one request and dropped when the response is sent. No webfonts, no analytics, no CDN. Because nothing is stored, there is nothing to delete and no breach surface — closing the tab is the entire cleanup.
| Problem | What happens to it |
|---|---|
| Duplicates | Exact matches on email and phone first, then fuzzy matching you approve group by group. Blank fields on the record you keep are filled in from the copies removed. |
| Dead email domains | Live DNS check that the domain publishes a mail server. No mailbox is ever probed — that gets your IP blocklisted and lies on catch-all domains. |
| Disposable & shared inboxes | 8,000+ throwaway providers, plus info@, sales@ and friends, flagged separately from real problems. |
| Misspelt domains | gmial.com, hotnail.com, yahooo.com — a correction is suggested, never applied silently. |
| Phone numbers | Parsed against every country's published numbering plan, normalised to E.164, with region, line type and carrier. Handles missing country codes, 00 prefixes, leading zeros, extensions and spreadsheet damage. |
| Names | shaun mcdonald → Shaun McDonald. MARY-ANNE VAN DER BERG → Mary-Anne van der Berg. A name you typed as DeShawn is left exactly as you typed it. |
| Companies | Pvt Ltd, Private Limited and Pvt. Ltd. all converge, so records that should deduplicate finally do. |
| Countries & states | usa, U.S.A. and United States of America become one thing. So do KA and karnataka. |
| Junk | N/A, asdf, xxx, keyboard mashing, placeholder emails and empty critical fields — flagged against the row, never quietly dropped. |
A report that overstates its confidence is worse than no report, so these are stated in the tool itself as well as here:
Python. Google's libphonenumber for phones, RapidFuzz for matching, RFC-correct email validation, pandas for the table. Runs as a CLI or a library.
FastAPI, fully stateless. Every request carries its own file; the server holds nothing between requests and writes nothing to disk.
React and Tailwind, built into the Python package so one process serves both. System fonts, no third-party requests.
332 tests, including a labelled calibration set for the duplicate scorer that asserts the lowest true match still outscores the highest false one.