Now · updated 27 October 2026
What I'm building
- Bus fleet: one morning of passengers counted and cross-checked on one bus; hand-correcting crowded door frames so the next retrain moves. Tyre-pressure sensors and per-tyre distance are the next layer.
- Roti classifier: retrained on the August–September line batches; 85% of rotis decided automatically at 99.7%, zero bad rotis passed.
- Tiffins ERP, a catering CRM + ERP built at Workfloor AI, going out to caterers in the UAE.
- The Day I Will Be Able To Talk, a short film I edited, is going out to festivals, Cannes first.
- First year of Data Science & AI at Masters' Union.
Build log
Dated as it happened, newest first. The dates come from the files, the commits and the reports, not from memory.
Case studies for the roti factory and the bus counter go up on this site: the review clip, the occupancy cross-check, the class montage, the numbers.
Bus counter: low-power retrain on both buses' footage, six minutes with no thermal warning. Built a local tool for hand-correcting crowded door frames; the first seven corrected frames change nothing measurable, so a few dozen are needed.
Roti classifier re-validated after the retrain: 84–85.5% of rotis auto-decided at 99.7–99.8%, zero bad rotis passed, one false reject in 38 on unseen September days.
Latest cut of The Day I Will Be Able To Talk exported, the ninth since the first assembly in August.
Audited 18 new roti batches from the factory's Drive: two duplicates, 2,640 new images, the scorched-board class dropped. Retrained.
Bus counter: cabin-camera runs on both buses for the two-camera cross-check and the empty-bus reset.
Bus 8090: 45 minutes of empty bus through the counter. Zero boardings, zero alightings, one spurious track. Its frames become hard negatives. Training job packaged for a free Colab GPU so the laptop stops cooking.
First full morning counted on bus 7687: 53 boardings, 38 alightings from the rear-door camera, with a review video that draws IN/OUT counters on the frames.
Bus counter starts: DVR footage from bus 7687, RT-DETR pseudo-labels on 816 frames, YOLO11s fine-tuned on them.
Wrote up the Roti Factory build on LinkedIn: four months from a big-name CRM to a system built from scratch.
AceCard: WhatsApp lead capture goes live. A greeting from a tapped card becomes a lead, and a customer's reply captures their number.
AceCard's storefront promoted to production at buy.theacecard.co: the card designer with ten foil designs, press-ready PDFs, Razorpay checkout and Delhivery fulfilment.
Tiffins ERP: weekly planning, per-site menus, yield and consumption-norm production sheets.
First assembly of The Day I Will Be Able To Talk, a short film by Arpit Beniwal that I'm editing.
First roti classifier trained on the July line batches. Reject bar moved from 0.20 to 0.15 after a pale but good roti was auto-rejected.
Pass bar moved from 0.80 to 0.95 after an uneven-thickness roti passed at P(Approved) = 0.92. Thickness barely shows top-down.
Tiffins ERP: full catalog import, spreadsheet upload, demo data, food-cost analytics.
First roti dataset batch from the factory line. The 6 July batch turned out to be office test shots; it never went into training.
The factory sends the first defect photos on WhatsApp: eleven kinds of bad roti. The classifier starts here.
Co-founded Workfloor AI. Technical lead at AceCard.
Roti Factory ERP: physical architecture and infrastructure SOP written down.
Started the Roti Factory CRM + ERP from scratch, straight after Class XII exams and a few weeks of cold-calling businesses to sell websites.
Founded Smildly AI: CRMs, AI receptionists and websites for law and professional-services firms.
Joule: after Deepak Padaki's and Ankur Warikoo's replies, dropped the whitepaper plan for a consumer stress test.
Cold-emailed Narayana Murthy and Ankur Warikoo about Joule. Murthy replied in 44 minutes and forwarded it to Catamaran.
Growithevent built, solo: register, pay, scan.