Building Bharat Talaash on AWS: Turning Scattered Missing-Person Records into Explainable Leads
How our team used a two-way report flow, multilingual input, semantic matching, scheduled monitoring, and two-sided consent to support safer missing-person searches.
A search problem hidden inside a data problem
When a person goes missing, families need speed. But a search is not one simple database query. A missing report may exist in one police system. A found person may be recorded by another police unit, a shelter, a hospital, an NGO, or a citizen. Names may be spelled in more than one way. Age may be an estimate. Clothing can change. A location can be a landmark instead of an address. A poster may be an image instead of searchable text.
The public numbers show the size of the task. NCRB Crime in India 2022 recorded 83,350 missing children and 442,572 missing adults during the year. Together that is 525,922 reports, or about 1,441 per day when divided across the year. These figures describe reports made during the year. They do not mean all 525,922 people remained missing. [D1]
Recent local reporting shows that the stream continues. An analysis of Maharashtra Police data reported 30,113 missing adults from January to May 2025, close to 200 each day in one state. [N1] In Nagpur, police-record data showed 737 unclaimed bodies over 30 months, including 286 classed as unnatural deaths. [N3] We must not assume that missing-person and unidentified-body data sets describe the same people. The safe conclusion is that investigators and families face large, changing record sets that are worth comparing carefully.
What we built
Bharat Talaash is a privacy-first matching and monitoring network. It supports both Missing to Found and Found to Missing searches. A family can submit a MissingCase and search indexed found-person records. A citizen, shelter, hospital, or NGO can submit a FoundReport and search user-submitted and imported missing-person records.
The system returns up to 10 candidate matches. It shows the score, source, and field-level reasons. A candidate is not an identity decision. It is a shorter list for a trained person, family, or police team to check.
The current matcher holds 14,554 indexed found-person records. Each report becomes a 384-dimension text embedding. We use the open-source all-MiniLM-L6-v2 model through Sentence Transformers. The Sentence-BERT paper explains how fixed sentence embeddings and cosine similarity can make semantic search much faster than comparing every pair with a full BERT model. [R1]
Why we did not use one AI score
A missing-person search is too sensitive for a single black-box number. Our base rank uses 80% semantic similarity and 20% structured matching. Structured fields include name, gender, age or age range, height, complexion, clothing, visible marks, location, and time. The score is then adjusted for data completeness. A direct gender conflict is filtered.
A strong candidate must pass several checks: final rank score at least 80, completeness at least 40, and structured score at least 60. These are workflow thresholds chosen for the prototype. They are not a scientific statement that two people are “80% the same”. We still need a labelled pilot data set to calibrate thresholds and measure false alerts.
Natural input, but an editable report
Families may not know which database fields to fill. They may speak naturally, mix Hindi and English, or only have a poster image. We used Amazon Transcribe for speech input and configured automatic language identification across 10 language codes. The transcript enters the same extraction flow as typed text.
For typed descriptions and posters, Amazon Bedrock extracts a structured draft. The prompt tells the model to leave uncertain fields empty. It must not infer identity, ethnicity, health, or other sensitive facts from a face. The user sees the draft, edits it, and confirms it before submission. Nothing uncertain is silently filed.
AWS architecture
Amazon Amplify hosts our Next.js frontend. Amazon Cognito handles sign-in. API Gateway gives the frontend a small modular API. Python Lambda functions create reports, launch search jobs, return dashboard data, update case status, record consent, and manage extraction jobs.
DynamoDB stores the core workflow records: User, MissingCase, FoundReport, and Match. A source-sync checkpoint stores ingestion progress. S3 stores private user media and official evidence files. Uploads are limited to 3 MB, stored with server-side encryption, and deleted after extraction when they are no longer needed.
The open-source matcher runs on an EC2 CPU instance. The backend reaches it through AWS Systems Manager Run Command. This avoids opening a public matching port. The matcher can be replaced later without changing the public API or the frontend.
Search that continues after the first day
A one-time “no match” should not end a case. New reports arrive later. EventBridge starts a monitoring Lambda every two days. The worker reads the last checkpoint and queries only records added after that time, with a small overlap to avoid missing a boundary record. It does not scan the full database every time.
For the hackathon prototype, one active case can receive seven free checks over 14 days. The user can mark the case as satisfied, not satisfied, stopped, matched, or closed. A stable pair ID ensures that the same missing and found report pair does not produce the same notification again.
Consent before connection
Missing-person data can be misused. We do not publish phone numbers, email addresses, or home addresses. When a strong candidate connects two user reports, Amazon SES sends two separate emails. Each side sees that a potential match exists and is asked for consent. The system can share contact details only after both sides agree.
The dashboard is also private to the signed-in user. It shows active cases, found reports, potential matches, closed cases, search count, monitoring state, and recent activity. Owner checks, rate limiting, redaction, private file paths, report fingerprints, and audit history add basic protection against mistakes and abuse.
What the numbers mean
Our current facts are system facts: 14,554 indexed found-person records, 384 dimensions per embedding, top 10 candidates, a recorded 150-of-150 embedding test, two-day monitoring, seven free prototype runs, a 3 MB upload limit, 30 API requests per user per minute, and 10 configured voice language codes.
Some reductions can be calculated but must be described correctly. A top-10 list cuts a 100-record manual review by 90%, or a 1,000-record review by 99%. It does not prove that recovery improves by 90% or 99%. A two-day automated check reduces the maximum wait compared with a weekly check from seven days to two, which is 71.4%. It does not prove that every case is solved 71.4% faster.
What we learned
The main lesson was that responsible AI is as much about workflow as model choice. A strong matcher is not useful if it leaks an address, repeats the same alert, hides the reason for a score, or turns a candidate into a false identity claim.
We also learned to design for incomplete data. A good system must say “unknown”, preserve the source, and show the user what needs review. It must work in both directions because the person who reports someone found may be the first person able to create the missing connection.
Next steps
Our next step is a small, reviewed pilot with a police unit, NGO, shelter, or child-care organisation. We will measure median time to the first useful lead, records opened before the known relevant candidate, Recall@10, Precision@10, duplicate alert rate, and privacy incidents.
We have identified a 121,418-record ZIPNET unidentified-dead-body archive as a possible future import pool, subject to lawful use, schema checks, and data quality review. We will also add source adapters for approved government, police, NGO, and shelter data. We will not scrape closed social platforms or bypass official API rules.
Face similarity may be explored later as one optional ranking signal. It will not confirm identity. NIST reports that face recognition error can vary with age, sex, race, and image quality. [R3] Any trial would need clear permission, demographic testing, threshold evaluation, and human review.
Closing
Bharat Talaash is not an automated investigator. It is a safer way to organise evidence, search both directions, keep active cases moving, and help people review the most useful records first. Our goal is simple: fewer missed connections, clearer evidence, and no contact sharing without consent.
Blog ending block to customise
AWS Region used: ap-southeast-2
Team: Divyash Yadav:- Backend development
Meet Pardeshi:- Frontend and Integration
Note: This prototype provides candidate ranking only. Contact local police or emergency services for an actual missing-person case.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article