SHOWCASE
AssistLive Answers for Sales Calls
ProblemI noticed on sales calls that reps get asked things they are not sure of, and sometimes even for basic questions they would say “I will get back to you” to the client, especially the ones that are new to the job. That leaves a bad impression, particularly when it is something they are expected to know. It is not entirely their fault: there is a lot of information, things keep getting updated, and someone else might have discussed something with the client earlier on a call or over email. So it takes a lot of time and effort to prepare.
ApproachA desktop app that listens to the live call and answers from the company’s own documents in a floating, always-on-top overlay. A model transcribes, another model picks out the genuine questions, retrieval pulls the matching passages from the indexed doc set, and a stronger model writes the answer with its sources. Meeting profiles scope it to the call type, and a rep voiceprint keeps the seller’s own speech out of the trigger path.
In practiceThe answer lands in about a second in a translucent window the rep can drag anywhere, and when the documents don’t cover the question it says so instead of inventing an answer.
North-star metricSeconds between the prospect finishing the question and a usable, sourced answer being on screen. That is what forced the overlay, because an answer the rep has to alt-tab for has already missed the moment.
Why desktop, not a tab The two things this needs, capturing the meeting’s audio and holding a window above the call, are exactly what a browser tab cannot do. Electron for the shell and a bundled Python service for capture, diarization and retrieval; the cost is an installer instead of a URL.
Designed to be wrong out loud Every answer is graded against what retrieval actually found, and a weak match is labelled and told to escalate rather than dressed up as an answer. On a live call a confident wrong answer is far more expensive than an admitted gap, so the guardrail is the feature.
QuizzlyAI Quiz Builder & Live Match
ProblemWhen I was preparing for exams, quizzing myself worked far better than reading the same notes again. The problem was making the quiz. The material sat across PDFs, slides and a few web pages, and turning it into questions took longer than the revision it was meant to help with. And when a few of us wanted to revise together, sharing one and seeing how everyone did was its own hassle.
ApproachPoint it at one or more sources (PDF, DOCX, TXT or a fetched page URL), then set the question count, difficulty, a focus area and which of nine types to draft (MCQ, true/false, fill-in-blank, short answer, numeric, cloze, poll, ordering, matching). Every question lands in an editor with its answer key and explanation. The quiz then ships as a link, QR or embed, or as a live host-controlled match players join by code.
In practiceGive it a PDF or a deck and it drafts the questions, or write them yourself. The same quiz then plays two ways: on your own, or live with a group. Content to a playable quiz in under a minute.
North-star metricTime from “I have this material” to a quiz someone can actually play. Everything else follows from that: ingesting a document rather than a blank form, drafting the explanation alongside the answer, and publishing to a link instead of an export.
Where AI stops Generation drafts questions; it never publishes them. Each item lands in the editor with its type, difficulty, points and explanation exposed for review, and the quiz keeps a record of which source documents produced it. Getting an answer key wrong costs far more than taking longer to produce it, so the human check is the point rather than a formality.
Two modes, one quiz The same quiz works asynchronously (share a link, play alone, scored against a pass mark) and live (host opens a lobby, players join by code from their phones, host controls the pace). Building both off one question model meant no separate “live version” to maintain.
QueueLessVirtual Queues for Any Venue
ProblemI noticed at banks and service counters that you either stand in the line to hold your place, or someone hands you a paper token and you spend the wait watching a monitor for your number. Either way you cannot walk away. Someone has to be there handing the tokens out, and you have to keep checking whether your turn has come.
ApproachA QR at the venue opens a mobile queue, so a phone does the job of both the paper token and the number board. No kiosk, and nothing to install on either side. A business signs itself up and is live in about a minute, and customers, counter staff and the admin each work from their own screen.
In practiceCustomers join from their own phone, watch their position move in real time, and get told when they are nearly up, so they can walk away and come back.
North-star metricMinutes a customer spends physically standing in line, not time-in-queue. Someone can wait forty minutes happily if they spend them elsewhere, which is why the status page leads with position and estimated wait rather than a countdown.
Keeping the queue honest An open join link can be spammed from anywhere. The venue QR rotates every ~30 seconds and the link cannot be forwarded, so joining requires physically being there; optional CAPTCHA and a manual-approval mode with visual PIN matching cover higher-risk venues.
Built for the whole floor, not one screen Five surfaces share one live queue: customer phone, counter tablet, waiting-room TV, operator desktop and a platform admin. Counters, staff accounts, categories, operating hours and an audit log make it usable by a team rather than a single person.
QuickShareEphemeral File Rooms
ProblemAt presentation sessions during my MBA, the first few minutes went into getting the deck onto the one PC wired to the projector: copying it off a USB drive, mailing it to the account already signed in there, or signing into our own mail to download it. Asking groups to submit beforehand only moved the work onto a TA, who had to collect the files and load them onto that machine. And anyone who spotted a mistake on the day, or during the session before their turn, was stuck with the version they had sent. (This seems unlikely to happen at a premier B-school, but surprisingly it used to happen more often than you would expect.)
Another pain was sharing files within the group, where everyone had to contribute something. That ran on WhatsApp, which was its own mess, with nothing organised and a new message for every new version.
Then of course there was Google Drive or some other cloud-based sharing, but that meant opening a link, sharing it, signing in, and so on, which took longer than an ephemeral task like that deserved.
So these cases, which seem trivial, occurred many times during the course, but were accepted the way they were.
ApproachA room-based handoff: a 6-character code or QR opens a shared space, guests upload up to 500 MB per file without an account, and the host keeps control through visibility toggles, room and panel passwords, kick, and a hard expiry.
In practiceFiles move in seconds with zero setup, and the room deletes itself and everything in it on expiry, so nothing lingers on a borrowed device.
VerifyEmail for ExcelCheck whether an email is real, inside an Excel cell
ProblemDuring outreach, sales and marketing sent cold emails, and every address had to be checked first to avoid the risk of the sending mailbox getting blocked. I had already built a small app for it: paste a list or upload an Excel or CSV file, and get the results back in the app or as a file. (Other paid services for the same thing were available, but they cost more and none of them fitted how the team actually worked.) I realised that since the team lived in Excel, it would be even better if the validation happened directly in Excel. So I made this Excel add-in.
ApproachA service sits between Excel and the verifier. Every install gets 50 free checks a day, and filling the formula down hundreds of rows goes out as a single request.
In practiceYou type it like any other function and it answers in the cell. Nothing to paste in and no account to make.
If you want to go above the free limit, you can pay a small amount: 500 checks at ₹100. Let me know.
Zero setup over bring your own keyMaking people paste an API key into Excel loses most of them, and I did not want to ship a shared one. The service holds the key and the add-in holds nothing secret.
A random install ID, not accounts No sign-up, but each install still gets its own 50 checks a day. A per-IP cap sits behind it so that minting fresh IDs cannot drain a paid API.
A pasted command, not a download A downloaded installer trips Windows SmartScreen and antivirus warnings that scare people off. One line of PowerShell avoids the downloaded-file flag, and a download is still there for anyone who prefers it.
Self-hosted first, then serverless The first version was a Windows service in Go with a SQLite cache and a control panel. Fine for one office and unusable for strangers, so it moved to a Cloudflare Worker that costs nothing to run.
TaskBoardWork and Results for Every Team
ProblemManagers chase status updates in chat and spreadsheets, and the real outcome of finished work never gets written down. When review time comes, nobody can show what a team actually achieved.
ApproachA task tracker where every task is closed with an outcome note and a measured result, which a manager can accept or send back. Organisations get their own space with several teams, role-based access (admin, manager, member, leadership) and a weekly digest.
In practiceEvery closed task leaves a measured result, so at review time a team can show the impact it had instead of the tickets it finished.
Vibe-coded over a weekend as one of my first vibe-coded apps just to try out AI-based coding. (Not a novel / differentiated product.) It is production-ready though, and small teams can actually use it, so let me know if you want it. It is free.
North-star metricShare of closed tasks that carry a measured result, which is why a task cannot be closed without an outcome note.
Key trade-off One free, shared platform over per-customer servers. Each organisation automatically gets its own small database, so data stays separate while hosting costs stay at zero.
Fairness by design The hard part was making a results tracker that does not turn into surveillance. There is no ranking of people by task count, no mood scores in reviews, and no impact categories that would make one kind of work look less valuable.
Every app is fully functional. If you hit a problem, do let me know what went wrong. Thanks in advance for your feedback.