Amazon · Summer 2026

Software Development Engineer Intern

An agentic dashboard that audits the appointment-matching service against downstream shipment signals, replacing manual data joins and log analysis and reducing investigations from roughly one hour to under one minute.

The problem

The appointment-matching service is an automated decision system. When a match conflicted with the shipment signals recorded after it, the teams responsible had to reconstruct why the system produced that outcome.

Each audit required manual joins across several disjoint datasets, followed by log analysis to trace the decision. A single investigation took roughly an hour, and its conclusion was only as reliable as the manual correlation behind it.

Approach

I built an agentic audit dashboard that resolves this decision lineage programmatically. It automates those joins, pairing each matching decision with the shipment signals that followed it, and returns a deterministic, evidence-backed explanation rather than an inference. An investigation that previously required manual joins and log analysis now completes as a single query in under one minute.

Key decisions and rationale

A transport-agnostic core with thin adapters. Query, reasoning and serialisation logic reside in a single core, with lightweight adapters for each access surface. Every surface therefore returns identical results, and a correction made once applies everywhere.

Tiered access enforced at the code boundary. The core exposes a backend-selectable read path over a shared dataset, with access partitioned into tiers. Enforcing those tiers in code rather than by convention isolates privileged reads and unbounded scan costs from general traffic.

A closed reasoning loop with deterministic verdicts. The tool assembles evidence, commits to a verdict, reconciles that verdict against observed downstream state, and incorporates each discrepancy into the next evaluation. Because teams act on its output, every explanation is traceable to the records that support it.

Impact

Investigation time fell from roughly one hour to under one minute per case, and each result is reproducible from its underlying records.

Reflection

Designing for consumers outside my own team shifted the primary engineering constraints from correctness alone to contract stability, observability and graceful degradation when an upstream dataset is incomplete. A tool that passes its tests and a tool another team depends on carry very different burdens of proof.

Amazon · Summer 2025

Software Development Engineer Intern

A full-stack diagnostics feature for Amazon Vendor Central that gives more than 10,000 vendors real-time visibility into purchase order eligibility, reducing redundant support tickets by 76 percent.

The problem

When a purchase order was marked ineligible, the reason was recoverable only internally. Vendors had no way to diagnose the issue themselves. Every question became a support ticket, and every ticket required an engineer to trace system state manually.

The vendor's need was not a faster support response. It was an answer available at the moment the question arose.

Approach

I designed and shipped a full-stack diagnostics feature for Vendor Central that surfaces each ineligibility determination at its source, in real time, to more than 10,000 vendors worldwide. Moving that information to the vendor converted a human-mediated support path into self-service.

Key decisions and rationale

Explainable states instead of generic failures. The typed React front end renders each ineligibility reason as a discrete, explainable state. Vendors can understand and act on the specific cause rather than receiving a generic error.

Data access modelled on the interface. The system is event-driven and serverless end to end: a Java service layer behind a managed API gateway, on-demand compute, and a key-value store. Its access patterns were designed around the read shapes the interface required, rather than the structure of the upstream data.

Zero-downtime release. The feature was deployed against live vendor traffic without service interruption.

Impact

Redundant support tickets decreased by 76 percent.

Reflection

The most effective support improvement was to remove the need for support. Surfacing each determination at its source resolved vendors' questions before they became tickets.

Nudgi · Behaviour Design and Mobile Application

Design Engineer

A messaging application in which a friend can send a gentle nudge on an unanswered conversation, designed to make replying the easiest next step rather than a source of guilt.

Overview

Nudgi addresses a pervasive but rarely examined behaviour: the avoidance of social micro-responsibilities, such as replying to a personal message, under conditions of cognitive overload. A reply typically takes less than a minute, yet people routinely defer these messages for days. Each unanswered conversation then becomes a persistent source of mental load and guilt.

We designed and built Nudgi to reduce that burden. A friend can send a light nudge on a pending conversation, and the application schedules self-nudges for the times a person is most likely to respond. The project progressed from a three-part diary study, through assumption testing, to a working mobile application, with usability testing at every stage of fidelity.

The problem

The project began with one question: why does a task that takes under a minute become one that people avoid for days? Existing research on task initiation, including the five-minute rule and the cost of context switching, explains the delay itself. Our focus was its psychological cost. Under cognitive overload, small social obligations accumulate into a continuous mental load, and the guilt grows the longer a reply is owed.

This reframed the objective. The product did not need to make people reply faster. It needed to make replying feel lighter.

The competitive review placed nine tools, from Habitica and Todoist to Notion and Jira, by task context and nudge frequency. Most target professional work, leaving personal communication comparatively underserved. Nudgi was positioned in that space, with nudges sent by friends rather than generated by a system.

Research approach

We conducted a three-part study: a pre-study interview, a five-day diary study in which participants logged the social micro-responsibilities they postponed, and a post-study interview. Participants were recruited through a screener survey.

A diary study was chosen over interviews alone because these tasks are frequent, brief and dependent on context. Such tasks are difficult to recall accurately after the fact. Logging each instance as it occurred captured the circumstances of a delay rather than a later reconstruction of it.

Key insights

Awareness alone changed behaviour. Several participants reported that keeping the diary made them more responsive before any intervention was introduced, and several realised the tasks were easier to complete than they had assumed. This suggested that a product could succeed by making pending replies visible, without applying pressure.

The strongest opportunities occurred at transition points. Journey mapping showed that participants were most receptive when they had finished one task and were deciding whether to begin another. This defined when a nudge should arrive, not only what it should say.

Responding felt like work. Participants described pending replies as obligations that weighed on their minds, particularly during periods of fatigue or stress. Many also reported guilt about the consequences of not responding.

Responsiveness depended on context. Whether a participant replied varied with the time of day, their schedule, the perceived intensity of the task, and whether the sender was a friend or a professional contact.

Testing the assumptions

Before committing to features, we mapped every assumption the concept depended on by importance and by strength of supporting evidence. We then tested those that were both important and unsupported.

The most consequential assumption was whether a nudge from a friend would be received as care or as pressure. The answer determined the tone of the entire product. Three related assumptions were prioritised for testing: that people want to become more responsive to friends, that they are willing to nudge friends and hold them accountable, and that attachment to a friend's avatar is strong enough to prompt a reply.

Design decisions and rationale

The resulting scope was deliberately narrow, and each feature responds to a specific research finding.

One-tap nudges within a conversation. A nudge is sent with a single action from inside the thread. Nudging a friend should never require more effort than the reply it encourages.

Self-nudges timed to likely response windows. Automatic nudges are scheduled for the times of day a person is most likely to reply, applying the transition-point insight directly. Users set the frequency, timing and quiet hours for these nudges.

A prioritised inbox. Nudged conversations move to the top of the inbox, and messages stay marked unread until the user replies. A separate waiting view lists friends who have not yet replied, ordered by how long they have been waiting. Visibility, which the diary study showed to be effective on its own, does most of the work.

Snooze and dismissal on every nudge. A single swipe answers, snoozes or clears any nudge, and users can let go of a message to remove it permanently. The snooze is not a convenience feature. It is the mechanism that prevents the system from becoming a new source of the guilt it was designed to reduce.

Close friends. Users can mark their closest people as close friends, so that the conversations that matter most are always prioritised.

Visual identity

Because the product reminds people of an obligation, its visual language carried much of the responsibility for feeling supportive rather than supervisory. The brand pairs Jua for display type with Nunito for interface text, and uses a soft palette of blue, butter yellow and blush pink on a cream ground in place of conventional alert colours. The style tile defined the intended tone as cheerful, supportive and positive.

Users select a penguin avatar during onboarding and customise its costume and background colour. This establishes a small sense of ownership before the application asks anything of the user.

Testing and outcome

Each stage of the design, from low to medium to high fidelity, was evaluated through usability testing. The tone of the nudge copy was revised more often than any interface element. In a product built on social accountability, language proved to be a core part of the interaction design rather than a finishing layer.

The final application demonstrates the complete loop: a conversation goes unanswered, a nudge arrives, and replying becomes the simplest available action.

Reflection

Nudgi showed that behaviour change does not always require stronger incentives. The most effective lever in this project was visibility, delivered with enough care that it did not register as pressure.

Roots · Cross-Platform Mobile Application

UX Researcher and App Developer

A mobile application of guided lessons, conversation games and a shared scrapbook, through which immigrant parents learn culturally relevant communication skills and practise them with their children.

Overview

Roots addresses intergenerational communication gaps in immigrant families, where parents and children often struggle to connect across cultural and generational distance. The application is built on the premise that this connection is a learnable skill rather than a fixed condition.

The application teaches a communication concept, provides a low-pressure activity through which parent and child practise it together, and preserves a record of the outcome. I served as UX researcher and app developer on the project.

Research

The team conducted nine interviews with parents, teenagers and a clinical expert. Participants were recruited deliberately across ages, cultural backgrounds and socioeconomic contexts, so that the findings would not reflect a single type of household.

Synthesis through empathy mapping revealed a consistent tension. Every participant wanted greater closeness, yet none had culturally specific tools for achieving it. Parents wanted to connect but did not know how to begin the conversation. Children wanted the same, and interpreted the resulting silence as disinterest.

This finding defined the product's purpose. The barrier was not motivation, which both generations already had, but the absence of a structured way to begin.

The pivot

The most consequential decision emerged from low-fidelity testing. The original concept treated parent and child as equal users of a single shared interface. Testing showed that their needs diverged significantly, and that the dual-user design weakened the product.

Parents engaged less readily when the experience also addressed their children. The shared interface also made vulnerable moments feel observed rather than safe.

The team therefore narrowed the design to a single primary user: the parent, who learns culturally relevant communication skills and then initiates practice with their child. This decision simplified every subsequent choice, from content to navigation, and turned a broad concept into a focused, usable product.

Trade-off. Narrowing to the parent removed the child's independent entry point into the application. The team accepted this because testing showed that the parent's willingness to engage was the condition on which the entire experience depended.

Technical decisions and rationale

React Native with Expo. I built the prototype in React Native with Expo so that a single codebase could run on both iOS and Android without maintaining native projects. This kept development effort focused on the experience under test.

Supabase as the backend. Supabase provides authentication, translated game content and lightweight session metadata, without the overhead of operating a custom server.

Functional internationalisation. Translation is handled through react-i18next, which made a working language toggle possible rather than a simulated one. For families in which parents and children are often most comfortable in different languages, this was a core requirement rather than an enhancement.

Randomised content within a fixed structure. Game content is stored remotely and randomised at retrieval, so that repeat sessions differ, while the lesson sequence remains fixed. Families can return to the same lesson without the practice becoming repetitive.

A deliberately scoped prototype. The prototype implements one complete path, lesson to game to scrapbook, to validate the interaction model before investing in a full curriculum engine. An onboarding survey indicates how the application could adapt to each household's context. The project documentation states explicitly which behaviours are fully implemented and which remain curated. Deciding what to build in full and what to stage was a deliberate part of the engineering work.

Next steps

The natural extensions of the prototype are the two components it deliberately staged: a curriculum engine supporting a full sequence of lessons, and onboarding that adapts content to each household's context.

Reflection

Roots reinforced that the most valuable product decision is often a reduction in scope. Defining a single primary user resolved questions that no amount of interface refinement could have answered.

Dosti Project · Stanford Indo-Pak Dosti Forum · Brand Identity and Growth

Co-Founder, Design and Growth Lead

The visual identity, merchandise and audience growth strategy for a student-founded Track II diplomacy initiative that promotes people-to-people connection between India and Pakistan.

The brief

Dosti Project is a 501(c)(3) nonprofit I co-founded to advance Track II diplomacy between India and Pakistan: sustained, unofficial exchange among students, artists, founders and former officials, conducted outside formal government channels. For more than seventy-five years, citizens of the two countries have lived as neighbours with limited ability to visit one another. Borders are frequently closed, visas are often refused, and most people encounter the other country primarily through news coverage that emphasises conflict.

The brand therefore had to introduce not a product with a clear benefit, but a relationship its audience had been conditioned to distrust.

Strategy

One principle shaped every subsequent design decision: lead with similarity, not difference. The political relationship between the two countries is complex, but their cultures share a great deal. They share a common language written in different scripts, a cuisine known by different names, and traditions in music, humour and urban life.

The identity needed to feel familiar to audiences in Lahore and Delhi simultaneously, without reducing either culture to the other. I deliberately avoided a neutral or diplomatic tone. Neutrality communicates distance, and distance was the problem the brand existed to address.

The wordmark

I translated this strategy into a single wordmark. Dosti means friendship and belongs equally to Urdu and Hindi, so the logo was designed to be read in three ways at once. Latin letterforms carry the English word, while their construction draws on the sweep of Nastaliq and the connecting headline of Devanagari.

The mark is intentionally one form rather than three translations placed side by side. This reflects the movement's central argument: these traditions were never wholly separate.

Audience growth

Alongside the identity, I led audience growth, building the movement from no existing presence to more than 15,000 followers across both countries and the diaspora. The content strategy followed directly from the brand strategy: publish the overlap. Posts featured shared vocabulary, shared recipes, shared humour and side-by-side images of cities in both countries.

I excluded political content entirely. It consistently provoked division in comment sections, whereas content built on recognition was widely shared. The highest-performing posts were those in which users from Karachi and Mumbai tagged one another under the same image.

Merchandise

I extended the identity to products designed for everyday use. One tote presents an illustrated inventory of objects both countries claim as their own: karak chai in its glass, a kite, a lotus, a dhol, a decorated rickshaw, a cricket bat and a mango. Another is styled as a dictionary entry, defining dosti through shared cultural references such as chai, butter chicken, doctor bano ge ya engineer? and aur sunao?

Each piece was designed so that the wearer identifies with shared experience rather than nationality.

Conference

The identity now supports the inaugural Stanford Indo-Pak Dosti Conference at the Hoover Institution. The conference convenes more than forty speakers across three continents, including former heads of government, founders, investors, filmmakers and artists, around people-led partnership. It sold more than 380 tickets and included a startup demo day with venture capital investors.

I extended the visual system across the conference website, the printed programme and the event environment.

Directions not taken. I developed and set aside two earlier website directions. The first used plum as a full-bleed background, which read as formal and institutional, closer to a diplomatic summit than a grassroots movement. The second framed a light page with the border motif along both edges; it was warmer, but it drew attention to the ornament rather than the speakers. The published site retains the motif as an accent and gives visual priority to the people featured.

Reflection

In a context shaped by distrust, the most persuasive design choice was restraint. Emphasising recognition over argument allowed the brand to reach audiences that political messaging would have lost.