Nobody asked me to.

Strategic Validity System

Boards commit to strategies constantly and rarely write down why. A market assumption quietly stops being true, nobody notices, and the strategy keeps running on a foundation that is already gone. Effectus Research needed a way to make that failure visible before it becomes expensive, so during my MBA internship there I built the software for it.

• Every commitment is scored on five weighted parts, the CARMR framework: what you are committing to, the assumptions it depends on, the reasoning that links them, whether your terms mean the same thing to everyone, and how it gets reviewed. Weighted 15/35/25/10/15, they roll into a single 0 to 100 Commitment Integrity Score.
• Disagreement is a feature, not noise. If someone on the team holds a minority position that contradicts the stated assumptions, it stays in the permanent record. It is never resolved by vote or quietly dropped.
• Language drift gets caught automatically. The same term can mean different things to different people in the same room. The system measures how far a term's stated definition has drifted from how it is actually used elsewhere in the record, and flags it before that ambiguity becomes a disagreement nobody can name.
• Nothing is overwritten. Every edit, activation, or review is saved as an immutable, versioned snapshot, so you can always see exactly what the team believed at any point in time, and what changed.

LiveCoaching

Most interview advice is generic and unmeasured. I wanted something that watched and listened to an actual practice session and told me, specifically, what happened: how fast I was talking, where my eyes went, what I said in the first minute that mattered most. So I built a Mac app that turns a rehearsal into evidence.

• Upload a job description and your CV and the app generates interview questions tailored to that specific role, not a generic bank.
• It watches and listens while you answer, tracking pace, pauses, filler words, eye line, and posture, camera use is optional and the app works fully without it.
• Every note ties back to something real, a line you actually said or a measurement actually taken, not a generic tip.
• Everything stays on the Mac. Once the app is set up, it runs with no internet connection at all.

The coaching is grounded in published research on first impressions and interview validity, not intuition, and I built the marketing site, a launch film, and a signed, invite-gated distribution build around it.

AceItChamp, with ShikshaSathi inside it

Almost every AI tutoring product launches in English first, which quietly excludes hundreds of millions of students in smaller Indian cities. I started with ShikshaSathi, a Hindi-first AI learning platform fine-tuned on the national curriculum, with an answer-validation framework so the AI was checked before a student ever relied on it. Ola backed the project. AceItChamp grew out of that work: an AI revision and assessment platform with adaptive study plans, and ShikshaSathi lives on today as its AI tutor.

• Piloted in three coaching institutes, with the design iterated weekly against real feedback from the students and teachers using it, alongside product analytics.
• A five-step learning loop: diagnostic assessment, a personalised path, interactive lessons with an AI tutor, real-time progress tracking, and achievement milestones.
• Built for patchy connectivity, with an offline mode that keeps downloaded lessons, basic quizzes, and progress tracking working without a connection, syncing once you are back online.

Virtual Maddy

LiveCoaching is built for a structured interview session. This is the version I keep running in the background for everything else: a small floating widget that sits in the Mac menu bar and quietly coaches whatever I am doing, not just rehearsed answers.

• It listens continuously, detecting when I am actually speaking versus silence, transcribing it, and a coaching thread turns that into live feedback on what I am saying, fed back to the floating widget in real time.
• Camera and screen are optional add-ons, each running its own capture thread that feeds a vision-coaching pass alongside the audio, so the widget can react to what it sees as well as what it hears.
• It behaves like a real menu-bar app, not a script window, complete with mic-permission handling, a settings dialog, and a session timer, and it stays out of the way until you need it.

AI Salary Predictor

Most salary tools give you a number with no reasoning behind it. I wanted the opposite: a predictor that tells you exactly which inputs moved the estimate, and by how much, the same discipline I want from any AI system with real consequences.

• A Random Forest model, trained on 200,000-plus records across nine factors: role, experience, education, skills, industry, company size, location, remote status, and certifications.
• Every prediction ships with its reasoning, using SHAP values to show exactly which inputs pushed the number up or down, plus a 95 per cent confidence range instead of a single false-precise figure.
• A live interface built on Next.js, with the model served through a separate FastAPI backend, so you can change an input and watch the prediction and its explanation update together.