Check whether your medicine's batch has been flagged as substandard or spurious by India's drug regulator — in seconds, from data that's otherwise buried in PDFs.
Every month, the Central Drugs Standard Control Organisation (CDSCO) tests samples of medicines sold in India and publishes a "Not of Standard Quality" (NSQ) alert — a list of specific drug batches found substandard, contaminated, mislabelled, or outright spurious (counterfeit). In a single recent month, 70 batches were flagged, from antibiotics to anaesthetics.
That information is life-critical and completely public — and almost no one can use it, because it ships as unindexed PDF tables on a government portal. No patient is going to read a 9-page PDF to check the strip in their hand. SafeDose turns those PDFs into a question anyone can ask: "Is this batch on the list?"
 |
 |
 |
The dataset is built by an ETL pipeline that reads the actual CDSCO NSQ alert PDFs and extracts every row:
S.No | Product/Drug Name | Batch No. | Mfg Date | Expiry | Manufactured By | NSQ Result | Tested By
backend/app/parser.py handles the real-world mess: it locates the header row (which repeats on every page), maps columns by fuzzy name (CDSCO's headers drift between months), pulls only true data rows, and classifies each free-text failure reason ("Dissolution (Buffer Stage)", "Particulate Matter", "Declared Spurious") into a coarse category. The raw source PDF is committed under backend/data/raw/ so the pipeline is fully reproducible:
python -m app.parser data/raw/NSQ_Alert_March-2025.pdf \
--period 2025-03 --source-url "https://cdsco.gov.in/.../March-2025.pdf" \
--out data/nsq_records.jsonThe shipped dataset is CDSCO's March 2025 alert (70 real flagged batches, 64 distinct medicines). Adding a month is one command — drop in the PDF, run the parser, re-seed.
A flag applies to a batch, not a brand, so batch number is the identity. backend/app/engine.py returns one of three states, and the design is deliberately conservative because the dangerous failure mode in a drug-safety tool is false reassurance:
| Result | When | Severity |
|---|---|---|
| Batch flagged | the exact (normalized) batch is on an NSQ list | 🔴 danger |
| Caution | a flagged batch shares the number but a different drug name (batch numbers aren't globally unique), or other batches of the same medicine are flagged | 🟠 caution |
| Not found | nothing matched — explicitly not a safety guarantee | 🟢 clear-but-hedged |
Two details that matter:
- Batch normalization —
AD-204,ad 204andAD204all compare equal (punctuation and case are stripped), because the number on the strip rarely matches the PDF's formatting exactly. - Collision honesty — a short batch like
0081can appear under unrelated drugs. An exact batch hit is only called a confident danger when the drug name is also similar (token-overlap + sequence similarity on the molecule, ignoring dosage-form noise and bracketed brand names). Otherwise it's surfaced as a caution with the matched record shown — never hidden, never overclaimed.
Example. A patient holds an Adrenaline injection, batch AD-204:
query: drug="Adrenaline", batch="AD-204"
→ normalized batch AD204 matches a flagged record
→ drug name "Adrenaline" is similar to "Adrenaline Bitartrate Injection IP (Reoline)"
→ BATCH FLAGGED · danger
"A government lab found this exact batch substandard. Do not use it."
Maker: M/s. Rathi Laboratories (Hindustan) Pvt. Ltd., Patna — Reason: Appearance / particulate matter
GET /api/check?drug=&batch= three-state result with the matched records and reasons
GET /api/alerts?q=&category=&spurious= browse/filter all flagged batches
GET /api/stats dataset summary
GET /api/meta disclaimer
GET /api/health
Consistent {"error": {code, message}} envelope; nothing about the query is stored.
docker compose up --build
# frontend → http://localhost:3100 API → http://localhost:8100/docsLocal development:
cd backend
python3 -m venv .venv && .venv/bin/pip install -r requirements-dev.txt
DATABASE_URL=sqlite:///dev.db .venv/bin/python -m scripts.seed
DATABASE_URL=sqlite:///dev.db .venv/bin/uvicorn app.main:app --reload --port 8100
# in another shell
cd frontend && npm install && npm run devcd backend && .venv/bin/python -m pytest # 32 testsThe suite runs the real March 2025 PDF through the parser (asserting known real entries are extracted), pins every engine state including the batch-collision and "not-found-is-not-safe" cases, and exercises the API against the real dataset.
- Data comes only from official CDSCO NSQ alert PDFs; the source URL and alert period travel with every record, and the raw PDF is committed for audit.
- Failure reasons are kept verbatim from CDSCO; the category is a derived convenience, never a substitute for the original text.
- The UI repeats on every surface: confirm on cdsco.gov.in, and "not found" is not a guarantee of safety.
- It's a snapshot. The shipped dataset is one month's alert. CDSCO publishes monthly; a production deployment would run the ingester on a schedule.
- CDSCO only catches what it samples. Absence from the list means "not caught," not "safe" — the tool says so, loudly.
- Name matching is heuristic. Indian brand names map to molecules in messy ways; the engine errs toward showing more (caution) rather than hiding a possible match.
- Automated monthly ingestion — a scheduled job that pulls each new CDSCO alert, diffs it, and notifies subscribers if a medicine they take is flagged.
- State drug-alert coverage — state FDAs publish their own NSQ lists; fold them in for far wider coverage.
- Barcode / 2D-code scanning — newer strips carry GS1 codes with batch + manufacturer; scan instead of type.
- Manufacturer risk view — repeat-offender manufacturers across months, to surface systemic bad actors.