There are two ways to use AI. Most people know one. I shipped fifteen systems with the other.
KEEP SCROLLING ↓ishmam@os %direct: take everything. plan it. build it.
Ask AI, get answers. Direct AI, get systems - you are scrolling one right now.
168,550+PEOPLE SERVED
ACROSS 15 SYSTEMS · ALL BUILT SOLO
NO ESTIMATES WITHOUT MATH · SOURCES: VERCEL ANALYTICS, ADMIN DASHBOARDS, USAGE LOGS, SESSION TRANSCRIPTS
Five carry full build logs - prompt zero onward, every quote preserved verbatim. Hover a row to peek inside.Tap a row to step inside.
NO.01HR, Finance, Tech and Product truth lived in five different sheets and ten systems - synthesis lived in two people's heads.
One internal screen over the company's real data - org, people, finance, product. Built solo, by the COO, with a coding agent.
402 people on one living org tree · 5 tools → 1 screen · ~80 min to ship a live revenue lens.
An IBA course - 73 on roster, ~69 actively taught - needed grading infrastructure for 3 sprints and 25 groups.
Built over a weekend: cap-and-median peer-rating math and a 40-30-20-10 rubric a spreadsheet cannot enforce.
145 grades published, 637 peer ratings across 2 review windows. Students used it all semester.
One dashboard, end to end - students, schools, referrals and registrations in a single view.
2,365 students · 1,595 school records · 3,709 referrals · 365 new registrations.
Every event card was pixel-pushed by hand - roughly 2 designer-minutes per card, at volume.
Feed it a list, get print-ready personalized cards back.
~1,500 cards generated ≈ ~50 design hours that simply stopped being work (1,500 × ~2 min/card).
An interactive map of Bangladesh's electoral geography, browsable district by district.
300 constituencies mapped.
Itineraries, phrases and budgets in my pocket - built for exactly one user and one trip.
No metric claimed. Lived on it for 12 days. Some systems are allowed to be personal.
The IBA alumni directory, live at amraibagraduate.com.
5,127 visitors · 11,803 page views in the last 12 months.
A personal-recap prototype the engineering team adopted.
~20,000 users generated their review at first launch - the Shikho Facebook group flooded with shares. The weekend prototype shipped inside the production Android app mostly intact, which is rare for a prototype handed to a real engineering team.
2021, pre-AI: a hand-coded Telegram bot serving COVID stats.
~150 users at peak, mostly Grameenphone colleagues who humored me.
2021: is Bangladesh under lockdown tomorrow?
Co-built with Muhammed Asif Khan, 15 months into learning to code. Launched around 30 Jun 2021.
Featured in The Daily Star, 11 Jul 2021.
Direction is the job. The output is rarely the point - the system that keeps producing the output is the point. Here is the loop, run on a real incident:
BETON KEMON · BLANK PAGES ON EVERY IPHONE · YOUR SCROLL RUNS THE LOOP
The loop above is one incident. This is that loop at volume - every Claude Code session on my laptop, across a two-month window.
I typed 3,566 instructions. The machine took 187,368 actions to carry them out. Direction, amplified 53 times.
378 SESSIONS · 31 PROJECTS · 772,412 WORDS TYPED (~9 NOVELS)
Not one long push - a rhythm. Every day of the week carries work, and the lightest of them still averages 38 instructions. The only two quiet days in the window were back-to-back.
Every project I run opens with a CLAUDE.md - my standing orders. Mine them all and the same base spine surfaces. That spine is my method.
Every claim on this page came out of a memory - mine. Below is a public mount of it: eight chapters, 15 builds, 176 mined sessions - queryable. No model answers you. It only replays what is on file - nothing invented, nothing generated.
SELF-LOGGED · MY.FLIGHTRADAR24.COM/ISHMAMCHOWDHURY · VISITS VERIFIED AGAINST TRAVEL RECORDS
I embed inside my own company like an engineer who ships, not a consultant who advises - and I can do the same inside yours: working systems, standards your team actually keeps, and people who can run them after I leave. This whole page is that portfolio.
I install the harness your org is missing: agent contracts, decision records, custom skills, enforcement hooks, eval loops. The same machinery that runs the systems on this page.
From the business problem to a working production AI system - guarded, verified, documented so it survives me leaving.
Your people build with AI on your real problems - the format delivered to LightCastle's consultants. Artifacts by end of day.
Plenty of people can demo AI. Far fewer have shipped production systems that real teams depend on - and then taught those teams to build the same way themselves.
