For investigators
Every person on file, scored against the case, with the reasons showing.
Lead Engine combines the records an agency already holds, weighs them with fixed and documented point values, and hands the investigator a ranked lead list, the evidence behind each point, and recommended next steps.
Ready for pilots. Built and tested on a fully synthetic city; real accuracy is measured on your own closed cases first. The figures on this page come from that synthetic test.
What it does
From a case description to a short list in minutes
Eight categories of evidence
Forensics (30 pts), vehicle (25), phone (25), physical description (20), history and MO (20), associations (15), geography (10) and confirmed calls or posts (15). Every point traces to a record.
Witness and victim statements
Each statement carries a confidence level and viewing conditions. A second witness strengthens shared details; a hesitant one counts for less.
Stolen cars and stolen cards
A plate seen near the scene that belongs to a victim is treated as a stolen getaway car, not a suspect. Card use after a theft becomes a time-and-place lead.
Rideshare and camera leads
Where no vehicle was seen, the engine suggests rideshare records by warrant and a camera canvass of nearby homes and businesses.
Suspect workup
For a known or suspected offender: supporting evidence, contradicting evidence, and the actions that would confirm or clear them, side by side.
Next-action plan
Prioritized steps for the case, with draft request language and the legal process each step needs (voluntary request, subpoena, warrant).
Jail-call check
For the top 10 leads, asks Call Review two questions: has this person been named in a confirmed jail call, and did any inmate call this person's phone after the offense? A call to their number is only a flag for a reviewer to listen; it earns no points until the reviewer confirms what was said.
Ready-made social media searches
Builds the Google, Facebook Marketplace, OfferUp and X searches a detective would type by hand (posts near the scene that night, the stolen items for sale, the getaway vehicle, the top leads' public pages) and opens each with one click in a normal browser. The engine does not monitor anyone or pull anything on its own; the detective looks and decides, and each search is written to the case log.
Scoring model
Fixed weights, open math
The model is additive and documented, so an investigator, a supervisor or a defense attorney can recompute any score by hand.
| Category | Max | What earns points | What reduces them |
|---|---|---|---|
| Forensics | 30 | A print or DNA hit, scored by where it was found; a firearms trace to a purchaser sits a tier below a print because guns change hands | A hit is a name to test, not a conclusion: every named person gets a workup listing what supports and what contradicts |
| Vehicle | 25 | Registered or associated vehicle matches make, color, body style; plate read near the scene in the time window | Plate read far from the scene; vehicle reported stolen (reassigned to the stolen-vehicle rule) |
| Phone | 25 | A number from the victim's phone that is new, spikes before the offense or was the last contact, resolved to a person and placed near the scene by records obtained lawfully | Requires corroboration, because a phone is often registered to someone else |
| Physical description | 20 | Sex, race, age within 3 years, height within 2 inches, marks or tattoos described by a witness | Contradictions are capped at −8; a sex mismatch demotes rather than removes |
| History and MO | 20 | Prior offenses of the same type or method, recent release, repeat pattern | Recency decay: a record half as old counts about twice as much |
| Geography | 10 | Home, work or known hangouts near the scene, graded by distance | On shift at a workplace more than a mile away excludes |
| Associations | 15 | Known associates already linked to the case; ties to the victim through prior reports, shared addresses or call history | Capped so that association alone never produces a high score |
| Calls and posts | 15 | A jail call or public post that a reviewer or investigator has confirmed refers to this case | Unconfirmed hits earn nothing; this category alone can never put someone on the list |
Hard exclusions: in custody at the time of the offense, GPS monitoring placing the person elsewhere, or a verified work shift more than a mile away. Minimum: a lead must have support from at least two categories before it is listed.
What it is, and what it is not
Lead Engine is data-driven lead identification. It is not predictive policing.
What Lead Engine is
- Data-driven lead identification for a crime that has been reported
- Ranks the people already on file against what the victim said and what the records show
- Shows the record behind every point, and the detective can open it
- Recommends next steps in the order the answers come back
- Proven on your own closed cases before anyone relies on it
What Lead Engine is not
- Not predictive policing. It starts with a reported crime and never forecasts one.
- Not probable cause. A rank is an investigative direction, never evidence.
- Not a black box. Fixed, written rules.
- Not surveillance. Department records and lawfully obtained results only; it collects nothing new.
Recommendations only. Department policy and the law control every step. The investigator decides.
Measured results
Our toughest test
A synthetic city of 500,000 generated residents, 844 closed cases with a known offender in each, built from published rates. Witnesses misjudge details at realistic rates, one offender in ten has no record at all, and lab results that come back after the suspect is named are not counted.
The engine also backtests itself: pointed at an agency's own closed cases, it reports how often the known offender landed in the top 5, calibrates what each score band means in practice, and can refit its weights on half the cases while holding the other half out to confirm the change helped.
A fairness check in every report: the engine ranks on evidence only. Afterward, the results are reviewed by race, sex and age against who was actually arrested, the same way stop data is reviewed, to show it doesn't lean toward any group.
Worked example · synthetic data
One case, start to finish
How a detective would work one case in Lead Engine, tab by tab. Fictional case, no real people or records.
Commercial burglary · Case 26-0417
2:10 a.m. forced entry at a phone store. A passing driver saw one man, late 20s to 30s, about 5'10", leave in a silver sedan. The store's card reader was taken; one of its stored customer cards was used at a gas station four miles away at 4:23 a.m. A plate reader on the corridor logged a silver Honda Accord at 1:58 a.m.
What the detective had on day one
The detective enters the description, the vehicle, the time window and the stolen-card details. Everything else the engine already has: records, booking descriptors, DMV, plate reads, custody status.
| Vehicle | 22 / 25 | Make, color and body match; plate read 0.3 mi from the scene inside the window |
| Physical | 14 / 20 | Sex, height and age within range; beard in 2025 booking photo vs "clean-shaven" (−2) |
| History and MO | 12 / 20 | Commercial burglary 2023, forced entry, overnight; counts less because it is three years old |
| Geography | 6 / 10 | Card used 0.6 mi from his listed address |
| Associations | 0 / 10 | None on file |
| Jail calls & posts | 0 / 15 | Nothing confirmed yet (see the next tab) |
Every person on file was scored the same way. A lead must have points in at least two categories to appear; a wrong detail lowers a score instead of removing the person.
What the engine checked, and what it left to people
Only what a person confirms (a reviewer for calls, the detective for posts) is added to the case and counted in the score.
Suggested next actions, in priority order
Each step names the legal process it needs and comes with draft request language. Suggestions, not orders: agency policy and the prosecutor decide.
How it ended in the simulation
The gas-station video showed the silver Accord and a bearded driver using the card. Lead A was the generated offender. The jail call to Lead B's phone, once reviewed, was unrelated and was dismissed; Lead B earned no points from it.
The engine's job ended at the ranked list and the plan. The detective's began there: the video, the lineup, the interview and the charging decision were all human work, with every point in the score traceable to a record a prosecutor or defense attorney can check.
How it works
Nothing new to collect. Everything connected.
Start from the data you already hold
Records, booking descriptors, DMV and stolen-vehicle reports, plate readers, custody status, supervision data, witness statements, and the jail calls and posts an investigator has confirmed. Each source feeds the categories it belongs to; nothing is bulk-loaded that a case does not call for.
Weigh eight kinds of evidence, in the open
Fixed, documented point values for forensics, vehicle, phone, physical match, history and MO, geography, associations and confirmed communications, adjusted for how old a record is and how confident a witness was. A wrong detail demotes a person instead of removing them. The weights can be refit on an agency's own closed cases.
A ranked list, the reasons, and the next move
Every lead shows which records earned which points and how often leads at that score turned out to be the offender. Each case comes with recommended next steps in priority order, draft request language, and the legal process each step requires. A score is a lead, never probable cause.
Questions we hear first
Where do the accuracy numbers come from?
From a fully synthetic city of 500,000 generated residents and 844 closed cases with a known offender in each. No real records were used. Real-world accuracy is measured on a partner agency's closed cases during a pilot.
Does this replace an investigator's judgment?
No. It ranks and explains; people decide. Every score is a lead that needs independent corroboration, and every next step is a suggestion that follows agency policy and legal review.
How does it avoid flagging the same people every time?
Leads need support from at least two categories, generic traits alone cannot produce a high score, prior-contact factors are capped and decay with age, and a fairness check in every report reviews the results by race, sex and age against who was actually arrested.
Where does it run, and who holds the data?
It is designed as an analytics layer inside the records, CAD and plate-reader platforms agencies already use, under the agency's existing CJIS-compliant hosting and access controls. Each office's data is isolated and never used to improve the system for anyone else.
Fit
An analytics layer, not another system
Lead Engine is designed to run inside the records, CAD and plate-reader platforms an agency already uses, under its existing CJIS-compliant hosting and access controls. It is offered to platform vendors under license and to agencies through those platforms.
- Inputs: RMS and booking records, DMV and NCIC stolen-vehicle data, ALPR reads, custody and supervision status, witness statements, confirmed communications, and by-warrant rideshare and card data.
- Outputs: ranked lead list with per-point evidence, suspect workups, next-action plan with draft request language, backtest and fairness reports.
- Every search, score and export is logged with the model version and assumed discoverable.