Work: eight jobs with measured results

A gold factory's production system

70 of 70 planned screens built, running on a permanent database since August 2026

The full job

What arrived: Metal left a central desk on a paper slip, went through casting, filing and polish, and came back lighter. The difference was explained from memory.
What we built: A production system where the weight arithmetic is done by the database itself, not by the screens, and every movement is written to a record that is never edited afterwards. Screens cover planning and production, the metal ledger, stock, outside work, refining, quality, dispatch, returns, purchase and reports, all in the trade's own words.
Measured: 70 of 70 planned screens built. 8 values the database calculates itself and 17 rules it enforces. 189 screen designs published for the owner to walk through before the build started. Running on a permanent database since August 2026.

Client name withheld

3 platforms web, Android and iOS, live from one codebase

The full job

What arrived: A legal AI company had a case law research product that had to reach advocates on a phone, a tablet and a browser. Features sat unfinished instead of reaching users, and nobody could say which version a lawyer in the field was running.
What we built: One app released to web, Android and iOS from a single codebase, with search over Indian judgments, AI search, judgment summaries and follow up questions. Before every release we open the finished app file and search it for the new code. That check found fixes that worked on the test copy but were missing from the live app, and got them released.
Measured: Live on web, Android and iOS from one codebase. The Android build of version 1.8.0 finished in 12 minutes 37 seconds and was checked to contain the new code. Tested on a real Android 14 phone. On a real account, 9 real judgments came back with follow up questions and four section summaries on desktop and mobile.

Client named with permission

27 of 29 written lines found in the raw footage of the reference episode, 93 percent

The full job

What arrived: A Marathi daily serial studio delivers six episodes a day. Raw footage arrives with no scene number, no take number and no clapperboard, from two cameras at once, and the editor first has to find which performance belongs to which written line.
What we built: A pipeline that takes a day of camera cards, writes down the Marathi dialogue, matches it to the shooting script, lines up the two cameras and assembles a first edit the editor can play. An editing app beside it shows the take under each line and lets the editor swap in another take. The editor stays in charge of choosing.
Measured: 27 of 29 written lines found in the raw footage of the reference episode, 93 percent. On two shoots it had never seen: 19 of 23 and 38 of 43 lines, 83 and 88 percent. Still weak, and written down: the edit has no silent reaction shots yet, which the editor's own cut uses a lot.

Client name withheld

51,993 documents loaded into seven browsable branches

The full job

What arrived: A labour law compliance product had search and a chat assistant on a database that looked full and was not. A reviewer said state laws were missing, overtime and minimum wages were blank and the section structure had collapsed. The question was whether to fix it or start again.
What we built: We checked each claim against the live data instead of arguing about it, then fixed what was really wrong. Missing law text was recovered from the source website's own document files, with a check that keeps scanned images out and lets only real text in.
Measured: 51,993 documents loaded into seven browsable branches. State legislation went from 0 documents with text to 240 across 32 states. Of the reviewer's claims, two were true and two were false: overtime held 45 documents with real text. The answer handed back was fix, not scrap, with the counts behind it.

Client name withheld

45 minutes to transcribe all 84 videos, about 42 times real time

The full job

What arrived: Paid video courses sat unwatched on a drive, because nobody has 31 hours. The useful part was a few dozen tactics with numbers, buried inside the videos.
What we built: A pipeline that pulls the audio out of each video, turns the speech into text, and writes short playbooks that a person or an AI assistant can search. It never deletes a video on a guess: it writes a list of what is safe to delete, and a person decides.
Measured: 84 videos and 31.3 hours, measured rather than taken from the drive label. All 84 transcribed in 45 minutes, about 42 times real time, with no broken or partial results. 12.06 GB of video reduced to 956 MB of audio first. The limit is written down: on large video files the speed drops, and one 116 MB file took 16 minutes.

Built in house

0 pieces lost across two real meetings recorded end to end

The full job

What arrived: Meeting note takers join the call as a guest, and since March 2026 the large platforms make the host admit them by hand every time. On a client's call, that means asking permission in front of the client.
What we built: A Chrome extension that records the meeting tab and the microphone on your own computer, so nothing joins the call. The recording is saved in small pieces as it goes, so a browser crash loses seconds and not the meeting, and a checker proves each recording is complete.
Measured: Two real meetings recorded end to end, 20 minutes 15 seconds and 13 minutes 18 seconds, with 0 pieces lost. The first transcript ran to 2,812 words of mixed Hindi and English with both sides heard. Still open: the written summary is not yet produced from real transcripts.

Built in house

30.1 percent of the stocks trading in early 2018 no longer trade, a gap the usual data hides

The full job

What arrived: Investors test share strategies in spreadsheets and stop at the first result that flatters them. The brief: type a strategy the way you would say it, and get an honest test over Nifty 50 history.
What we built: A web app that turns the sentence into an exact set of rules you can see and edit, shows every assumption before anything runs, then tests it over Nifty 50 price history. Splits, bonus issues, dividends and similar company actions are adjusted for.
Measured: Live at backtest.kantabai.life. 504 of 1,676 stocks trading in early 2018 no longer trade, a 30.1 percent gap that the usual data hides, and honest data lowers returns in every period. A strategy that looked like 21 percent a year came out no better than the roughly 10 percent index return on honest data, and the fair number is the one shown.

Built in house

2 sites live from one codebase, India and the UAE

The full job

What arrived: A technology services firm needed a corporate site, and then the same site again for its UAE company under a different brand, without paying to build it twice.
What we built: One codebase, two sites. The second brand changed only the logo, the page titles, the site name and the founder line. Both sites sit on the client's own hosting, not ours.
Measured: Both are live. The UAE phone number was changed in all six places in the code and checked on the live site. Still weak, and written down: both sites build each page inside the visitor's browser, so Google first sees an empty page.

Client named with permission