Dhruv Singhal · Product Manager & Builder
Product Intern, Growth · The Sleep Company (Jul–Oct 2026)
E-commerce and AI products — stakeholder input, user research, and product specs turned into decisions teams can build.
Four case studies below. Each states the decision, the trade-off, what was tested, and what is still unproven.
AI trust and evaluation · focused scope · D2C retention definitions and privacy · rider-to-rider matching and fare fairness.
AI
Photograph a sick plant, get a diagnosis you can act on.
Decision: Require a source link on every recommendation, and judge prompt changes against a fixed golden set, not by feel.
Co-built with Dilpreet Grover · product, eval, orchestration
SourcedA source link required on every recommendation · offline eval result withheld; field accuracy not measured
Product
Your task list, pinned behind every window on the desktop.
Decision: Pin the widget behind every window, not on top of them, and refuse attention: accountless, local-first, no dock icon.
Solo · product, design and engineering
0Failures · release gate run, 15 Sep 2026
Self-hosted A/B testing whose statistics engine is allowed to say keep running: a sequential verdict instead of a bare p-value, and assignment that puts the same user in the same variant every time. Built solo; evidence so far is a synthetic demo tenant, not production use.
Solo · product and architectureLocal build · not deployedEvidence: BuiltA personal analytics desk for my own Zerodha account, running on my own laptop: technical indicators, Monte Carlo VaR, ensemble forecasting, a NIFTY 500 screener and a FIFO trade journal. Private by design; the tests run on synthetic data only.
Solo · personal tool, product and architecturePrivateEvidence: BuiltBuilt so a candidate can prepare for the TCS NQT without a paid prep course: 404 practice questions and 10 previous papers (counts per the repo README), quizzes, flashcards and progress tracking. Installable as a PWA and usable offline.
Built the web app · question bank started by three other contributorsLiveEvidence: BuiltDesigned to help solo founders find a co-founder they would actually build with: swipe-based matching shaped around founder-discovery pain points. A two-person build; usage was never instrumented, so there is no user number to report.
Two-person build · I owned productArchivedEvidence: Self-reportedHelps a long-distance couple stay in sync when they can't read the room in person. Cycle phase, mood and stress drive daily guidance for her, him and us. Private, used by two people. Privacy is enforced at the database rather than the UI: 575 RLS policy checks, offline-first, hosted on free tiers ($0 a month, self-recorded).
Solo · product and engineeringPrivateEvidence: TestedTreated my own portfolio as a product: spec and success criteria before code, then versioned iterations from v1 (Jan 2026) to v6 (Sept 2026), with the decision trail written up as its own case study.
Solo · product, content and buildLiveEvidence: TestedEach number says where it stops. Case-study and article counts are coverage, not outcomes.
A prompt is an artefact. An eval harness is a product. Here's how I scope LLM features so the output doesn't surprise users — or me.
There are two ways for a PM to get technical fluency wrong: not seeing why a “simple” feature takes three sprints, and reaching so far into the code that you rewrite an engineer’s PR. Where the useful middle sits.
I ranked my next side project on five criteria. The AI-heavy ideas scored highest on excitement and lowest on everything that predicts a finished tool.
Two products can share a DAU line and have opposite futures. A retention curve cut by acquisition cohort tells you which one is actually compounding.
When a design review ends with “it looks off”, nobody can pick anything up. The method I use to turn that verdict into ranked, located, owned tracker rows.
About a year of internships across three companies, and no vision decks. What the job actually is at intern level, and five lessons from getting it wrong.
Then we should probably talk. Email is the fastest way to reach me; LinkedIn works too. More on how I work is on the about page.
