Five days after the Cockroach Janta Party called off its weeks-long sit-in at Jantar Mantar, Delhi Police floated a ₹15.66 crore tender for 338 cameras fitted with facial recognition, placed along the very roads that link Parliament, Jantar Mantar, Ramlila Maidan and the India Gate hexagon. Two and a half months later, the police de-notified Jantar Mantar as a protest site and named Ramlila Maidan as one of its replacements - a ground already covered, twice over, in the camera list filed in July.
A tender built around "VIP/VVIP" routes
The tender, floated on the Government e-Marketplace on 30 July 2026 and still carrying the file name "DRAFT RFP 338 CCTV Final," seeks 248 fixed cameras, 43 pan-tilt-zoom units and 47 automatic number-plate readers, each wired for facial recognition, audio capture and "people gathering" detection. Sites named in the document include the VVIP gate of Ramlila Maidan, Turkman Gate, Delhi Gate, Rajghat, Mandi House, Patiala House and every approach to the India Gate circle. The winning vendor will run the system for five years, with footage stored for a minimum of 30 days. Officially, the justification is crowd management for Republic Day and Independence Day. In practice, the geography overlaps almost exactly with the arteries used by protest marches through central Delhi.
Oversight that exists on paper but not in practice
This is not an isolated purchase. It follows the same template used for a 194-camera tender in North District earlier this year, and sits alongside a separate deployment of over 6,000 cameras across fifty police-station areas, roughly a tenth of them already equipped with face-matching software. None of this spending appears to be drawn from the city's larger Safe City programme; it is being assembled district by district, contract by contract, with no public framework governing how faces are matched, against which database, or for how long the resulting data is kept. The only retention figure in the tender concerns disk storage, not personal data. There is no mention of the Supreme Court's privacy jurisprudence, no data protection impact assessment, and no reference to the very Standing Order the Home Ministry has cited to Parliament as the legal basis for facial recognition at protests - a document that, notably, never uses the words "facial recognition" at all.
Errors that suggest haste, not rigour
The tender's technical annexures appear to have been copied from the earlier North District bid, complete with mismatched page numbers. Camera quantities for pan-tilt-zoom and number-plate units are transposed in the bill of quantities. The video-management software specification lifts language directly from a single manufacturer's marketing material, a practice that GeM's own procurement rules are meant to prohibit. One set of coordinates places a camera roughly 40 kilometres from where it is supposed to stand. The integrity pact meant to deter corruption names monitors whose terms expired years ago, with no reference to the vigilance official appointed to oversee large contracts since 2025. Taken individually, these might be dismissed as clerical slips. Taken together, across a document authorising facial recognition on every camera in a public square, they point to a procurement moving faster than anyone is checking it.
Why the location matters
Facial recognition deployed on fixed infrastructure differs from a one-off police sweep: it is persistent, automated and indifferent to whether the crowd below is a parade or a protest. Civil liberties groups have already flagged the Jantar Mantar deployment to the Supreme Court, where a judicial committee is examining the legality of surveilling demonstrators. Moving the designated protest site to Ramlila Maidan does not remove that camera from the picture - it was listed in the same tender nine weeks before the order was signed. Whether that is coincidence or design remains unproven. What is documented, in the tender's own pages, is that the ground chosen for future protests was ground already marked for facial recognition, microphones and crowd-detection software, procured through a bid too error-ridden to say with confidence what, exactly, the city is buying.