Adamma Infoservices needed candidate identity verification that could run at volume without a human comparing photographs one pair at a time. We specified, built and delivered it in three months. It has processed over a thousand verifications since January 2026.
verifications processed
A background verification firm receives three photographs of a candidate from three points in the hiring process - the government ID, the interview, and the reference image on file. The question is simple to ask and slow to answer at volume: is this the same person in all three?
Done manually, it is a bottleneck that scales linearly with hiring volume. Adamma needed it automated, consistent, and defensible enough to stand behind in front of their corporate clients.
They also wanted a second signal - whether a candidate's face appears elsewhere on the open web, under a different name or a different professional history than the one declared.
All three images are ingested, faces detected and aligned, and every pair compared for identity. The output is a determination across all three sources rather than a single check - so a mismatch is localised. An ID that does not match the interview photograph is a different finding from a reference image that matches neither, and each points the verifier somewhere different.
Built on AWS Rekognition for detection and matching, with AWS Bedrock in the pipeline.
The reference image is dispatched through SerpAPI for reverse image search, with results retrieved, filtered and captured as evidence on the candidate's verification record.
Getting this to focus on the person took engineering. General image search anchors on whatever is most visually distinctive in a frame - in a professional headshot, that is the suit and the spectacles, not the face. We implemented an automated face-cropping stage that isolates the face before dispatch, stripping clothing and background from the query so the search works on the subject rather than their wardrobe.
The system is embedded in Adamma's day-to-day verification workflow rather than running as a trial. Steady four-figure throughput over seven months is the measure that matters: it survived contact with real operations.
Face detection accuracy in live operation
From kickoff to production delivery
Candidate verifications processed
Of continuous uninterrupted operation
A verification system in hiring sits between a company and someone's job offer. We designed the operating model accordingly.
The system operates entirely within Adamma's consented background verification process, with candidate consent obtained at onboarding. Biometric data is handled under India's DPDP Act 2023, and system output is advisory to a human verifier at every stage.
Identity verification is not a technology problem alone. It is a question of where the machine stops and a person takes responsibility.
How we designed the operating model
Start with one department and one term, on your own rubric.
