Anuti
Shrivastava.

Most B2B reorders still arrive as emails and get retyped into the ERP by sales admins. It is slow, error-prone, and the intent behind the order never gets captured in structured form.

Most B2B reorders still arrive as emails and get retyped into the ERP by sales admins. It is slow, error-prone, and the intent behind the order never gets captured in structured form.

Built a seller-side reorder agent end to end on Next.js, Supabase and the Claude API. It reads an inbound email, identifies the buyer, resolves SKUs against contract prices and stock, and either submits the order, asks the buyer to clarify, or routes it for human approval. Deterministic guardrails on spend limits, credit and delivery feasibility gate every submission, every step is written to an audit trail, and a 14-case eval suite runs the agent in shadow mode before any prompt, model or policy change ships.

Built a seller-side reorder agent end to end on Next.js, Supabase and the Claude API. It reads an inbound email, identifies the buyer, resolves SKUs against contract prices and stock, and either submits the order, asks the buyer to clarify, or routes it for human approval. Deterministic guardrails on spend limits, credit and delivery feasibility gate every submission, every step is written to an audit trail, and a 14-case eval suite runs the agent in shadow mode before any prompt, model or policy change ships.

Working prototype with six pre-run scenarios covering every outcome, from straight-through order to hard stop. The live link is a locked exhibit (execution disabled). Contact me for a full walkthrough.

Working prototype with six pre-run scenarios covering every outcome, from straight-through order to hard stop. The live link is a locked exhibit (execution disabled). Contact me for a full walkthrough.

Early-learning apps are passive and one-size-fits-all. Built a voice-first AI tutor that adapts to a child and stays grounded in a real curriculum.

Early-learning apps for young children are passive and one-size-fits-all. The goal: a voice-first tutor that adapts to a child's responses while staying grounded in a real curriculum.

Full-stack voice tutor on Node.js with OpenAI Whisper, GPT-4, and OpenAI TTS. Built a FAISS-backed RAG pipeline grounded in the UK FS2 curriculum, with prompt engineering and adaptive question logic to prevent hallucinations. Owned end-to-end: product, architecture, RAG, frontend.

Built a full-stack voice tutor on Node.js with OpenAI Whisper for speech-to-text, GPT-4 for tutoring logic, and OpenAI TTS for character voices. Grounded responses in the UK FS2 curriculum via a FAISS-backed RAG pipeline to avoid hallucinations, with adaptive question generation and hinting logic. Owned end-to-end: product, architecture, backend, RAG, and frontend.

Working prototype validating voice-first, curriculum-grounded AI tutoring for early-age learners.

Working prototype validating that voice-first, curriculum-grounded AI tutoring is feasible for early-age learners.

Wonderloom needed a website that conveys an AI-personalised gifting product clearly enough to convert hesitant parents and gift buyers.

Wonderloom needed a consumer-facing site that explains an AI-personalised gifting product clearly enough to convert hesitant parents and gift buyers.

Shipped a responsive Next.js 14 site with TypeScript, Tailwind, and Framer Motion. Built a custom pastel design system, reusable book-cover and how-it-works components, deployed on Vercel with preview branches. Owned product direction, brand, UI, and full GitHub-to-Vercel pipeline.

Shipped a responsive Next.js 14 site with TypeScript, Tailwind, and Framer Motion. Built a custom pastel design system, reusable book-cover and how-it-works components, and deployed on Vercel with preview branches. Owned product direction, content flow, UI, brand positioning, and the full GitHub-to-Vercel pipeline.

Live consumer storefront supporting customer discovery and stakeholder demos for the AI storytelling product.

Live storefront supporting customer discovery, recruiter visibility, and stakeholder demos for the AI storytelling product.

Fitness apps tell women what they did, not how their body responded. Fitness OS is an intelligence layer that turns workouts into self-awareness, weeks into clarity, and months into real optimisation, designed first for the highest-churn segment: postpartum.

Fitness apps tell women what they did, not how their body responded. Fitness OS is an intelligence layer that turns workouts into self-awareness, weeks into clarity, and months into real optimisation, designed first for the highest-churn segment: postpartum.

Mobile-first React + TypeScript on Vite, Tailwind for styling, Recharts for charts, Framer Motion for motion. Transparent Recovery Score from sleep, soreness, and prior-intensity signals. AI insights run as deterministic correlations on seeded data so the demo behaves identically every time. Owned full stack: vision, PRD, design system, scoring logic, build, and deploy.

Mobile-first React + TypeScript app on Vite, styled with Tailwind, charts on Recharts, motion via Framer Motion. Hybrid design pulling Apple Health's clarity and Whoop's data discipline, with a transparent Recovery Score built from sleep, soreness, and prior-intensity signals. All AI insights run as deterministic correlations on 42 days of seeded data, so the demo behaves identically every time. Owned full stack end-to-end: positioning, four-pillar product vision, PRD, design system, scoring logic, build, and deploy.

Working prototype proving a postpartum-first intelligence layer can sit on top of an existing fitness platform without rebuilding workouts.

Working prototype proving a postpartum-first intelligence layer can sit on top of an existing fitness platform without rebuilding workouts.