Because the record must not forget.
A Service to Mankind initiative — stmorg.in
Sakshi — Sanskrit for witness — is an open-source, fully static public-accountability platform that tracks publicly reported sexual-assault cases across India. It is aggregated daily from public judicial records (eCourts / NJDG, High Court portals, Indian Kanoon) and credible media (RSS of established outlets), and presented on a fast, filterable dashboard hosted on GitHub Pages. Its single purpose is civic accountability: to make the scale, geography, and judicial status of these cases permanently visible, jurisdiction by jurisdiction — irrespective of anyone's identity, community, or politics.
Sakshi is a civic-accountability record of publicly reported cases, compiled from public judicial records and credible media. Counts reflect what was reported and recorded in public sources. It is not a substitute for official crime statistics such as NCRB data — reporting rates vary across regions and over time.
An accused person is presumed innocent until proven guilty. Any status shown reflects public court records as of the last update. Acquittals and quashed cases are shown with equal prominence.
| Service | Number / Link |
|---|---|
| National Women's Helpline | 181 |
| Police | 112 |
| Cyber Crime | 1930 |
| NCW complaint portal | ncwapps.nic.in |
These are shown on every page of the site as well. If you or someone you know is in immediate danger, contact the police on 112.
PUBLIC SOURCES
eCourts / NJDG · High Court portals · Indian Kanoon
NCRB · PIB / police press · RSS (established media)
│
▼
┌─────────┐ RawDocument(url, publisher, fetched_at, text)
│ FETCH │ httpx + asyncio · robots.txt · <=1 req / 2s per host
└─────────┘ honest User-Agent · ETag/Last-Modified · backoff
│
▼
┌────────────────┐ Gemini (gemini-2.5-flash), response_schema-
│ GEMINI EXTRACT │ constrained JSON over ALREADY-PUBLIC text only.
└────────────────┘ victim := null enforced · confidence required
│
▼
┌──────────┐ regex + NER PII stripping on EVERY string field.
│ SANITIZE │ Last gate before disk. 100% branch coverage.
└──────────┘
│
▼
┌────────┐ exact on CNR/FIR · fuzzy on district/date±3d/sections/court
│ DEDUPE │ CASE-ANCHORED (never victim) · ambiguous -> data/_review/
└────────┘
│
▼
┌────────────────┐ jsonschema every record · assign deterministic IDs
│ VALIDATE/SHARD │ write shards + summary.json + index.json
└────────────────┘ pii_guard as final assertion · atomic writes
│
▼
┌────────┐ data/summary.json · data/index.json
│ data/ │ data/{YYYY}/{STATE}.json (regenerated idempotently)
└────────┘
│
▼
┌──────────────────────────────┐
│ STATIC SITE on GitHub Pages │ Vite + vanilla JS · reads JSON shards
└──────────────────────────────┘
These are not preferences. They are encoded as automation, not merely prose.
Victim names, photos, addresses, family-member names, school / workplace, ages (beyond
the boolean minor), and any re-identifying detail are never written to disk,
committed, logged, placed in model outputs, cached, or kept in git history.
- Legal basis: Section 72 of the Bharatiya Nyaya Sanhita 2023 (formerly IPC 228A) criminalizes disclosing the identity of victims of sexual offences; Section 23 of the POCSO Act 2012 does the same for minors, extending to any identifying detail.
- The extraction prompt always forces
"victim": null, and a post-extraction sanitizer strips any PII-shaped field regardless. For minors, a record stores only: state, district, year, offence category, and judicial status.
Cases are matched and deduplicated on FIR number, court case number (CNR), police station, district, date, and court — never on victim identity.
Accused names are stored only when present in an official court record (judgment, order,
or cause list) — never from media alone. Media-only sourcing yields
"name_public_court_record": null and the label "Withheld (media-sourced, not yet in
court record)". Every case carries the persistent banner: "An accused person is
presumed innocent until proven guilty. Status shown reflects public court records as of
the last update." Removal-on-request is supported for acquitted / quashed cases; see
TAKEDOWN.md.
Every data point carries sources[] — {url, publisher/court, retrieved date}. Nothing
without a citable public source is published. Extractions with confidence < 0.8 are
quarantined to data/_review/ (excluded from Pages) for human review and are never
auto-published.
# 1. Clone
git clone https://github.com/ServiceToMankind/sakshi.git
cd sakshi
# 2. Create a Python 3.12 environment
python3.12 -m venv .venv
source .venv/bin/activate
# (or, with uv)
# uv venv --python 3.12 && source .venv/bin/activate
# 3. Install dependencies
pip install -e '.[dev]'
# (or) uv pip install -e '.[dev]'
# 4. Configure secrets
cp .env.example .env
# then edit .env and set GEMINI_API_KEY=... (never commit .env)
# 5. Run the full quality gate
make check
# 6. Run the site locally
cd site
npm install
npm run devmake check runs ruff, mypy --strict, pytest (>=85% pipeline coverage, 100% on
sanitize / pii_guard), jsonschema validation of all shards, pii_guard, the
summary.json size assertion, eslint, and prettier. Lighthouse CI runs on the
built site via make lighthouse (and in the deploy pipeline).
sakshi/
├── README.md
├── CONTRIBUTING.md
├── TAKEDOWN.md
├── SECURITY.md
├── CODE_OF_CONDUCT.md
├── LICENSE # MIT — applies to code
├── LICENSE-DATA # ODbL v1.0 — applies to everything under data/
├── .env.example
├── Makefile
├── pyproject.toml
├── schemas/
│ ├── case.schema.json # single source of truth for record shape
│ └── extraction.schema.json # pre-sanitized Gemini response schema
├── pipeline/ # Python 3.12, fully typed
│ ├── pii_constants.py # canonical forbidden-field list + PII regexes
│ ├── sanitize.py # PII stripping — last gate before disk
│ ├── validate.py # jsonschema validation + summary-size gate
│ ├── dedupe.py # case-anchored dedup + merge policy
│ ├── shard.py # validate, assign IDs, write shards
│ ├── extract/gemini.py # response_schema-constrained extraction
│ └── sources/ # one module per source -> RawDocument
├── scripts/pii_guard.py # final assertion over written data
├── data/ # generated — never hand-edited
│ ├── summary.json # < 50 KB (CI-asserted)
│ ├── index.json # shard manifest
│ ├── _review/ # low-confidence / ambiguous (not published)
│ └── {YYYY}/{STATE}.json # full records, sorted by date desc
├── site/ # Vite + vanilla JS static frontend
└── .github/workflows/ # ci.yml, daily update action
Humans never hand-edit anything under data/. The whole tree is regenerated
idempotently on each run, so re-runs are safe.
- Code — MIT License. See LICENSE.
- Data (everything under
data/) — Open Database License (ODbL) v1.0. See LICENSE-DATA.
- CONTRIBUTING.md — how to contribute, and the things that must never
be done (hand-editing
data/, storing PII in fixtures, adding frameworks to the static site, bypassingsanitize.py, looseningadditionalProperties:false). - TAKEDOWN.md — removal-on-request process for acquitted / quashed cases.
- SECURITY.md — responsible disclosure.
- CODE_OF_CONDUCT.md — community expectations.
Sakshi is a witness, not a judge. It records what public sources say, cites them, and lets the record stand.