Abel drafts police reports from body camera footage. The officer signs the report, and a DA may rely on it in court. This shows an evaluation layer that checks every drafted sentence against the footage and dispatch log, flags what isn't backed up, and learns from every officer edit.
Fictional example. The incident, people, times and numbers are made up to show the method. This is an outside concept, not Abel's product or data, and not affiliated with Abel.
Built for Abel · the short version
Every drafted report has to hold up in court, and today nobody can show that it does.
In this sample, 5 of 11 drafted sentences weren't backed by the footage or dispatch log. All 5 were caught before signing.
The data is made up. To make it real, I'd need a few hundred drafts paired with the reports officers actually signed, plus the footage timestamps they cite.
Who built it
Anirudh Devella, data scientist. Before this, in production:
8.3% forecast error at Bluelight (YC)
$14M+ in invoices benchmarked at Bluelight
AI Coach for 10,000+ athletes at Brandlete
150K leads scored at CoverRight
“I’d hire him again.” Alex Dixon, Co-founder, Bluelight (YC), managed Anirudh directly. Read it on LinkedIn
The eval layer sits between the draft and the officer
Abel already turns footage into a draft. The eval layer splits that draft into single claims, looks for each one in the evidence, and hands the officer a draft with every weak spot marked. Every officer edit then flows back as a labeled error.
02 · What the officer sees
A draft with the evidence behind each sentence
Click any sentence to see what the eval layer found. Green is backed by the footage or the dispatch log. Amber means nothing in the evidence backs it. Red means the evidence says something different.
Backed by evidenceNot in the evidenceEvidence disagrees
Incident report · Draft generated from body camera
Incident Retail theft
Case 26-031477
Location 412 Elm St, Elm Street Market
Date 03/14/2026
Reporting officer Ofc. J. Rivera, Unit 214
Evidence 24 min body cam, 7 dispatch entries
Officer signature: pending review
Incident timeline
Where the evidence for each sentence sits. Dispatch entries are above the line, body camera moments below. Sentence 9 has no marker because nothing backs it.
03 · Officer edits become data
Every correction is a free label
When an officer fixes a sentence, the change is tagged by error type and logged. Over thousands of reports this shows exactly where the model goes wrong, and which fixes to ship first.
↓ tagged and counted across every report
Error log, last 30 days
Illustrative counts across all departments. The model team works this list from the top.
04 · What the DA sees
A report that carries its own proof
EVIDENCE CHECKED
Evidence check · Case 26-031477
10 of 10
statements matched to footage or dispatch at signing
Checked against 24 min of body camera and 7 dispatch entries
5 statements corrected or removed by the officer before signing
Every correction logged with the evidence that triggered it
Today a DA has to trust that a drafted report matches the video. With the check attached, each statement points to the second of footage or the dispatch entry behind it.
The same numbers, rolled up by department, are what a chief needs to approve Abel, and what Abel needs to win the next department.
Next step
My first week with your real data
Diff drafts against signed reports and tag every officer edit by error type.
Rank error types by how often they still reach a signed report.
Ship the first plain-rule checks (times, addresses, quotes) for the top three.