Deepak Kaushish Sharma Email  /  LinkedIn  /  GitHub  / 

I build what moves the numbers. Marketplaces, travel, quick commerce, banking. Now with AI.

My work sits where product, operations and technology meet, at Foodpanda, OYO, Zepto, Dell and now Citi. The industries change; my curiosity doesn't. I find the lever that moves the metric, and I build it. Foodpanda's conversion went from 10% to 34%. 100K+ merchants went live at over 99% quality. OYO saved ~$10M. Along the way I've led teams of up to 100 across Europe, the Middle East, APAC and India.

AI has widened what's possible, and I was building with it early. At Citi I created an AI tool that saves analysts hundreds of hours. On my own I've shipped four AI products, from child development to film financing.

Agentic AI / RAG / LLMs
Guardrails / Evals / Governance
Python / React / Next.js
Node.js / AWS / MongoDB

Products I've Built

4 shipped, 4 live
Kidsightful01

Child development intelligence that gives parents real clarity.

A structured, age-specific assessment produces a personalized developmental profile: cognitive patterns, emotional tendencies, social dynamics, and learning style.

The engine analyzes responses against developmental benchmarks for ages 2–18 — insights a child psychologist would take multiple sessions to surface, from toddler milestones to adolescent patterns.

Next.js · LLMs · Clerkkidsightful.com ↗
Kritheon02

Behavioral intelligence that replaces market research with prediction.

Describe your product, campaign, or strategy and get a complete prediction report: how your audience will respond, what drives their decisions, and where they'll push back.

Independent behavioral profiles report purchase intent, emotional response, and reasoning — with statistical validation, confidence intervals, and segment breakdowns. Minutes, not 6–8 agency weeks. Validated across India, US, Canada, UAE, and UK.

Python · LLMs · MongoDB · Vercelkritheon.com ↗
CinePredict03

A virtual test screening for your script.

Upload a screenplay, get a complete greenlight package: audience intelligence, box office modeling, marketing direction, and talent impact analysis in one report.

The financial engine runs 10,000 Monte Carlo simulations with break-even and ROI at multiple budget levels — plus trailer hooks, taglines, viral moments, and risk flags. Native to both Hollywood and Indian cinema.

React · Python · LLMs · Three.jscinepredict.com ↗
InstaRecs04

A recommendation engine that makes everyday decisions effortless.

Tell it what you're deciding and get a personalized recommendation backed by reasoning — food, gifts, fashion, entertainment, travel — tailored to your context and constraints.

The engine learns your taste profile across categories and sharpens over time. Specific picks with explanations, not generic top-10 lists — handling budget, dietary needs, occasion, and mood.

Python · LLMs · Flask · MongoDB · AWSinstarecs.com ↗

Writing

Article + working code
Who Checks the AI That Checks the AI?Oct 2026

I gave an AI agent AI judges to check its work, then tried to break it.

An AI agent should never act on its own say-so. In this harness the agent only proposes. Deterministic rules, independent judges from different model families, and people for the uncertain cases decide what happens. A tamper-evident flight recorder keeps the evidence.

Run live on one laptop, the agent followed a prompt injection hidden in a support ticket, promised a discount nobody approved, and tried to push around a blocked branch. The checks caught every one. On 50 held-out decisions nothing unsafe got through, even with one judge deliberately compromised.

Python · Gemma 4 · d3-flash · Signed audit logRead the article ↗Code & demo ↗

Hobby Projects

Side builds
Nebula01

A storage intelligence desktop app that shows exactly where your disk space goes.

An interactive treemap of any folder or drive, byte-accurate duplicate detection (size → 128 KB fingerprint → full SHA-1), perceptual matching for similar photos, folder-vs-drive compare, and one-click undoable auto-organize.

Scans are indexed so reopening is instant and re-analysis only reads changed files — 120k files in ~80 ms. Nothing is ever permanently deleted: removed copies go to a restorable, drive-local trash.

Electron · Node.js · macOS · Windowsgithub.com/deepirex/nebula ↗
Page Curl02

Magazine-style page turns for React Native, with a real paper curl.

The page wraps around a cylinder that follows your finger, its back shows faint mirrored print, the page beneath is shaded, and a stack of sheet edges shows how much is left.

Works for fixed pages like a PDF rendered to images, and for flowing content: a pager lays reports and articles out as pages for every screen size and text size. Powers the reader in Kidsightful.

React Native · Skia · ReanimatedPrivate repo

How I Build

Method
01Full-stackPython backends, React/Next.js frontends, cloud-native on AWS.
02Real AI architectureLLMs, behavioral modeling, predictive systems. Not wrappers.
03ShippedAll four products are built, deployed, and running in production.
04SoloDesigned, built, and deployed end-to-end. No team, no outsourcing.
05Safe by designGuardrails, grounding, human escalation and an audit trail, so AI can be trusted with real work.

Career

Selected roles
CitiVP Technology, Digital Wallets & Agentic CommerceWorking on digital wallet integration and solutioning, biometric checkout authentication, and agentic commerce. Built an internal AI tool that drafts complete feature documents, saving analysts hundreds of hours; now rolling out across the bank.
ZeptoHead of CatalogLed a 40+ person catalog team; used LLMs to re-classify the catalog into an L1–L3 hierarchy and fix data quality at India's fastest-growing quick-commerce company.
OYODirector, Product & Content ExperiencesLed ~30 people; built an ML image-scoring model that lifted conversion 30%. Launched 700 cloud kitchens across 20 brands; ~$10M in direct savings.
Delivery Hero / FoodpandaDirector & Global Head, Content & Product OpsCarved out an international unit of ~100 people across Europe, APAC, Canada and Australia and ran it at 15%+ monthly profitability. Conversion 10% → 34%; 100K+ merchants onboarded at >99% quality.
DellEMEA Content Operations ManagerEight years serving Europe, the Middle East and Africa: online conversion +15%, CSAT +50%, repeat calls cut from 28% to 12%.

Let's build
something real.

deepak.kaushishs@gmail.com LinkedIn ↗