kshitij chauhan Get in touch
AI & Product
Delhi, --:--

Shipping products.

American ExpressBusiness Analyst
FMS DelhiMBA
Samsung R&DResearch Intern
IIIT Naya RaipurB.Tech, Computer Science
Chapter 1 of 4
IIIT Naya Raipur, B.Tech Computer Science

Rank 1, CS.
9.56/10.

Institute Gold Medal for Transformational Leadership. Institute Silver Medal for the highest GPA.

Samsung R&D Institute India

PRISM Excellence Award.

On-device machine learning inside Chromium. Keyword extraction 25% faster.

American Express

AI in production.

22 months. Complaint AI piloted in two countries. Semi-finalist at Innovate 2025.

FMS Delhi, MBA

CAT 99.58 percentile.

MBA, Faculty of Management Studies, University of Delhi.

2020Scroll
WorkPoint at a nameSwipe
205
complaint themes classified
78%
weighted accuracy, UK and Canada
500K+
complaints categorised
60%
less analysis time
American ExpressBusiness Analyst, 2024 to 2026
01

TAGS

In pilot, UK and CanadaAI that reads customer complaints and names the issueCase study
02

Dudebot

Innovate 2025, global top 5An AI first-day buddy for every new colleagueCase study
03

Audit trail

Passed auditThe complaint model that shows its evidenceCase study
Samsung R&DResearch Intern, 2022 to 2023
04

PRISM

PRISM Excellence AwardMachine learning that runs on the device, inside ChromiumCase study
VentureTechnology and finance lead, 2023
05

Samvaad

Finalist, Startup Expo, IIM AmritsarSign language to text and speech, and backCase study
Nights and weekendsOwn builds and research
Running live
Game theory, Next.js

Conflux

A lab for cooperation. Change the incentives, watch trust spread or collapse.

Open it liveNotesCode
DeepJS626s
First-Fit680s
Tetris689s
Time to finish every job
Reinforcement learning

DeepJS

A job scheduler that learns. Beats two standard heuristics on the benchmark.

NotesCode
AR planetarium
Published in IEEE Transactions

Learning games

Four games that teach teenagers to think like programmers.

Case studyPaper
Form: goodPose tracking
Computer vision

FitZen

An AI trainer that checks yoga poses and gym form.

Code
THREAD
Guess the word
Hack Reddit 2025, TypeScript

Hanging By a Thread

A hangman game built for Reddit.

Code
Policy number
Incident date
Claim amount
Documents
Filling the claim form
Google GenAI Hackathon

Insurance agent

An AI agent that explains insurance and fills claim forms.

Code
AchievementsBeyond the four chapters
1st of 426Constitutional Conclave 2024IIM Lucknow, national
1st of 400Logical reasoning quizCognizance 2022, IIT Roorkee
WinnerMachine learning track, AI Fitness CoachIndustry Academia Meet 2023, 100+ teams
Top 1%Elite certificate, Organizational BehaviorNPTEL, IIT Hyderabad, 2023
ICPC RegionalistCertificate of Achievement2023, among 1,000+ teams
ReviewerFrontiers in EducationAcademic peer review, since 2025
Taught150+ beginners, as lead teaching assistantScores rose from 45% to 70%, IIIT Naya Raipur
LedTech for the Technovate festival, team of 61,000+ sign-ups with zero downtime
WinnerChessAarambh 5.0, 2023
MVPCricket, Player of the TournamentAarambh 5.0, wicketkeeper-batsman
PublishedIEEE Transactions on Education, 2024A journal that accepts 15 to 20% of submissions
Certifiedin strategy and AIBCG, Google, Anthropic, 2026
Off the clock

Credits rolled.

Marvel's Spider-Man Remastered, finished October 2026.

Also: chess, cricket, air-rifle shooting, horse riding, geopolitics
Build
AI
NLP, zero-shot classification, RAG, agents
ML
Reinforcement learning, computer vision, SHAP
Code
Python, C++, SQL, TypeScript
Data and cloud
PySpark, BigQuery, GCP, AWS
Apps
React, Node.js, Next.js, Unity
Ship
Discover
User research, problem framing, storyboards
Decide
ROI models, business cases
Align
Leadership buy‑in, pitches
Deliver
Pilots, rollout plans
Measure
Success metrics, audit readiness
ResearchAuthor and reviewer
Research wall01 / 04
AuthorSSRN preprint, accepted at AIVR 2026

When to Act, When to Ask

When should an enterprise AI agent act, and when should it ask a person?

Read on SSRN
Act/Ask
AuthorIEEE Transactions on Education, 2024

A gamified learning framework to cultivate critical thinking skills in students

Four games, 50 high-school students, measured before and after.

Read on IEEE Xplore
11citations on Google Scholar
AuthorSpringer, 2023

Benchmarking of semantic segmentation enabled human activity recognition methods

A comparison of leading methods for segmenting the human body in images.

Read on Springer
walking
headtorsoarmleg
ReviewerFrontiers in Education, since 2025

Peer reviewer for Frontiers in Education

Reviewing research manuscripts for the journal.

Visit the journal
ReviewFrontiers in Education
LiveStraight from Medium, LinkedIn and GitHub
Medium connecting
LinkedIn connecting
GitHub connecting
All repositories

Want to connect?

kshitij7chauhan@gmail.com
Kshitij Kumar Singh Chauhan
LinkedInGitHubScholarMedium
Back to top
← All work
American ExpressIn pilot, UK and Canada

TAGS

Analysts were reading customer complaints one by one to work out what they were about. I built the product that does it for them, across 205 themes and two countries.

Built at American Express. Code and data stay on company systems.
My role
Project lead: framing, modelling, build, rollout
Where
American Express, customer experience analytics
When
2024 to 2026
Built with
Python, PySpark, BigQuery, React, Plotly Dash
60%
less time spent on thematic analysis
78%
weighted accuracy on the top 15 themes, in two markets
10+
teams serving themselves from the dashboards on top

01The problem

Complaints arrive as free text. To act on them, a business has to know what they are about.

A billing error, a declined card and a rewards dispute need different owners. That sorting was manual, slow and different from one reviewer to the next.

Two things made it harder. One complaint often covers several issues, so a single label loses information. And in a regulated business, a label nobody can explain is a label nobody can use.

02The user

Two groups depended on this work.

PrimaryAnalysts and reviewersThey read complaints and assign themes. They need speed and consistency, and they need to trust the label.
SecondaryCX and product ownersThey decide what to fix first. They need a ranked list of issues by volume and trend, without waiting for a report.

03What failed first

The obvious approach was to let the data organise itself. It didn't work.

Approaches tried, and the verdict

ApproachThe ideaWhat happened
ClusteringGroup similar complaints automaticallyGroups didn't match categories the business acts ondropped
EmbeddingsLet meaning, not keywords, define similarityDifferent issues sat close togetherdropped
Topic modelsDiscover themes from word patternsTopics blurred into one anotherdropped
Expert taxonomy + zero-shotExperts name the themes, the model scores themShipped, piloted in two marketsshipped

The root cause was the same each time: complaint vocabulary overlaps across themes. Customers use the same words (charge, card, call, refund) for very different problems.

The lesson: the taxonomy is a product decision. No model was going to discover it for us.

04The bet

  • Experts define the themes. The taxonomy grew from about 50 themes to 205, with an explicit Other bucket so the model is never forced into a wrong answer.
  • Zero-shot classification. A BART-MNLI model scores each complaint against each theme description. Adding a theme means writing a sentence, with no retraining.
  • Multiple labels per complaint. A complaint gets every theme it clears the bar for.

How a complaint moves through TAGS

Complaint textFree text, any market
205 theme descriptionsWritten by experts
Zero-shot scoringEach complaint against each theme
Multi-label outputEvery theme above threshold, or Other
Self-serve dashboardThemes by volume and trend

An invented example, to show the output shape. This is not customer data.

"I was charged twice for the same hotel booking and nobody on the phone could tell me when I'd get my money back."

Duplicate chargeRefund timelineCall-centre resolutionCard declinedRewards dispute

05Shipping it

ConstraintNo GPUsI parallelised inference so the full volume could be scored on the hardware we had.
AdoptionSelf-serve, not reportsReact and Plotly Dash dashboards replaced static reports. 10+ teams adopted them and reporting load fell 25%.
SpeedDeployment pipelinesStructured release pipelines for models and dashboards cut release time 30%.
ScaleTwo marketsI won leadership approval for a pilot in the UK and Canada, then ran it.

06What I took from it

  • Abandoning three techniques early saved months. Killing an approach is a product skill.
  • An Other bucket is a feature. A model that can say "none of these" earns more trust than one that always answers.
  • Explainability decided adoption. People used the system because they could see why it said what it said.
← All work
American ExpressInnovate 2025, global top 5

Dudebot

New joiners have a hundred small questions and are afraid to ask most of them. Dudebot is an always-on assistant that answers with sources, pitched with the business case to fund it.

Open it livePrototype codeAct-Ask paper
My role
Agent prototype, business case, problem framing, architecture, rollout plan
Format
Team entry, enterprise-wide contest
When
American Express Innovate 2025
Outcome
Global semi-finalist, top 5 by popular vote
Top 5
worldwide by popular vote, in an enterprise-wide contest
$2.89M
net annual savings at the cautious end of the model
1.1M
HR and IT queries a year across 75,000 colleagues

01The problem

The pitch opened with a storyboard about two colleagues.

Pam has just joined. Jim has been here for years. This is their week.

DAY ONE

What if I sound dumb? I’d rather fumble through than ask.

7:40 PM, SAME DAY

I wish I had a 24/7 buddy to answer the silliest of my questions.

DAY THREE

I can’t find the portal to mark my leaves. I searched the intranet and still can’t figure out the right link.

MEANWHILE, JIM

Why is scheduling across time zones so hard? I just need one slot that works for everyone!

FRIDAY

Is there anyone here who shares my interests?

ENTERS DUDEBOT…

Hello, Pam!

Ask me anything related to Amex. I am your dude!

1 / 6
Friction 1Afraid to ask

02The product

A retrieval-based assistant that lives where people already work and shows its sources.

  • Scope first. The pilot covers 300 to 500 high-value documents across HR, IT and onboarding, chosen by how often the questions come up.
  • A stated bar. 80 to 85% accuracy on the most repeated queries before any wider rollout.
  • It does things too. Scheduling a meeting or finding colleagues with shared interests takes one request.

Architecture

Where people ask
  • Slack
  • Company intranet
  • Outlook
Dudebot core
  • Conversation layer
  • Language understanding
  • Context and session manager
  • Orchestration engine
  • Security and compliance: sign-in, role-based access, audit, privacy
What it connects to
  • IT service desk
  • Knowledge base
  • HR and learning systems
  • Calendar

Every request passes the security layer before it reaches a connected system.

03The business case

A good idea still needs a number. I built the model from public benchmarks and halved every saving.

Time saved rarely turns fully into money saved, so only 50% counts. The pitch range was $2.89M to $7.52M a year, net of running costs. Move the assumptions and watch it change.

Adjust the assumptions

Queries per year (75,000 people × 15)1,125,000
Handled by Dudebot
Staff hours freed
Year-one cost, build plus upkeep
Months to pay back
Net savings per year

Upkeep is 22.5% of build cost a year. Sources: American Express Annual Report 2023 (workforce), MetricNet and HDI (ticket volume and cost), Forrester Total Economic Impact studies (HR queries, handling time), Gartner (automatable share), US Bureau of Labor Statistics (labour cost).

04The rollout

  1. Discover. Find the highest-friction support areas, align stakeholders, confirm access to knowledge sources.
  2. Design. Define the architecture and plan secure access to content.
  3. Build. Conversation interface, knowledge connection, automated workflows, structured content.
  4. Test. Functionality and security testing, then stakeholder approval. Not approved means back to build.
  5. Pilot. A small group, inside their daily tools. Measure, collect feedback, iterate.
  6. Roll out. Training, early adopters as advocates, then monitoring and new use cases.

05What happened

Dudebot reached the global semi-finals of American Express Innovate 2025 and finished in the top 5 by popular vote.

Building it raised a question the prototype couldn't answer: how much should an agent do before it checks with a person? That became Act-Ask, a research paper accepted at AIVR 2026.

← All work
American ExpressPassed audit

Audit trail

An in-house AI model in a regulated business has to explain itself before it can launch. I made explainability the feature that got this one through audit.

Built at American Express. Code and data stay on company systems.
My role
Built the pipeline and the explainability layer
Where
American Express, complaint capture
When
January to June 2024
Built with
Python, supervised NLP, SHAP
500K+
customer complaints categorised, replacing manual review
60%
higher review efficiency after automating the workflow
10%
fewer misclassified complaints

01The problem

Reviewers were checking complaint capture by hand, case by case.

A model could do the sorting. But this was a regulated setting, and an accurate model was not enough. Audit needed to see why each decision was made before it would sign off.

02Research

I shadowed the second-line review teams while they audited live cases. Watching the work showed where reviewers lost time and which mistakes mattered most to them.

Those frontline pain points became the list of classifier upgrades.

03The product decision

Treat the explanation as part of the output, not an afterthought.

  • Top 5 keywords per decision. Each output is attributed to the words that drove it, using SHAP.
  • Top 10 evidence points. Every decision traces back to the evidence behind it.
  • A ranked issue list. Open-ended feedback becomes a list teams can prioritise from.

What a reviewer sees. An invented example, not customer data.

"I was charged twice for one booking and have had no refund after three calls to the helpline."

Complaint: yesDriven by: charged twice, no refund, three calls

04Outcome

  • The model secured audit sign-off.
  • Review efficiency rose 60% and misclassification fell 10%.
  • CX and analytics teams used the ranked issue list to steer fix priorities.
  • The team later scaled this foundation into its multi-label workflow. That became TAGS.
← All work
Samsung R&DPRISM Excellence Award

PRISM

A smart device can't wait on a server to understand a command. At Samsung R&D I built machine learning that runs on the device itself, inside the browser engine.

Research internship. Code stays with Samsung.
My role
Research Intern, applied machine learning
Where
Samsung R&D Institute India
When
Sep 2022 to Jun 2023
Built with
Chromium source, Linux, Bash
25%
faster keyword extraction on the device
10
months inside a distributed research team
Award
Certificate of Excellence, Samsung PRISM

01The problem

An IoT device has little memory, little power and no time to wait.

The research problem was predicting devices and commands for IoT. Doing that on the device means models and text processing small and fast enough to live inside the browser engine.

02What I built

PlatformIoT features in ChromiumExtended the Chromium source code, tuned for data communication and on-device performance.
ModelsOn-device classifiersIntegrated several classifiers with an on-device parser to extract data and run inference efficiently.
SpeedKeyword extractionA lightweight pipeline in Linux and Bash for low-latency text processing.
TeamProduction-aligned modulesDelivered with the systems team across a remote, distributed group.

How the pieces fit, simplified

Page contentInside the browser
On-device parserExtracts the data
Keyword extraction25% faster
ClassifiersRun on the device
PredictionDevice and command

03Outcome

  • Keyword extraction ran 25% faster.
  • Pairing the classifiers with a lightweight parser improved on-device inference efficiency.
  • Samsung awarded a Certificate of Excellence for the work on IoT device and command prediction.

It was my first lesson in building models for a hard constraint, which is most of what shipping AI turns out to be.

← All work
VentureFinalist, Startup Expo, IIM Amritsar

Samvaad

More than 6 million deaf people in India rely on a language most hearing people can't read. Samvaad translates Indian Sign Language to text and speech, and back.

My role
Owned the technology and finance verticals
Type
Venture, with the Wadhwani Foundation
When
2023
Outcome
Finalist, Startup Expo, IIM Amritsar

01The problem

A conversation needs two directions. Most sign-language tools only offer one.

A deaf person signing to a hearing person needs the signs turned into words. The reply then needs to come back in a form they can follow. Break either direction and the conversation stops.

02The product

I designed the technology as a two-way flow.

Two-way translation

Indian Sign LanguageSigned to camera
TextShown on screen
SpeechSpoken aloud
SpeechFrom the hearing person
TextShown on screen
Indian Sign LanguageReturned to the signer

03The venture

We built Samvaad as a venture with the Wadhwani Foundation. I owned two verticals: the technology and the finances.

We pitched it at the Startup Expo of E-Summit 2023 at IIM Amritsar and reached the finals.

It was my first time carrying both the build and the business of the same product.

← All work
Personal buildLive

Conflux

Why do people cooperate when cheating pays better? I built a sandbox to watch the answer change as you change the rules.

Open it liveCodeEssay
My role
Solo: model, simulation, interface, writing
Type
Independent project
When
2025
Built with
Next.js 14, TypeScript, Tailwind

01The idea

Governance debates are arguments about incentives, held without any way to test them.

Conflux turns those arguments into something you can run. It models people as agents on a network. Each one chooses to cooperate or defect, earns a payoff, and learns from neighbours.

MechanismsTax, reward, punishmentSwitch each on or off and tune its threshold and rate.
NetworksWell-mixed or small-worldCompare how the same rules play out on different structures.
AgentsFour kinds of peopleEgoists, conditional cooperators, punishers and altruists.
ExperimentsSweeps and heatmapsRun factorial experiments, overlay results, export to CSV.

Noise is built in, because real people make mistakes and misread each other.

02Try a small one

This is a compact relative of Conflux, the classic spatial prisoner's dilemma. Every square plays its eight neighbours, then copies whichever neighbour scored best.

CooperatingDefecting

–cooperating, after 0 rounds

Push the reward past 1.8. Cooperation survives in clusters long after it should have collapsed. Structure protects trust.

03Why it matters to product work

Product decisions are incentive designs. A referral scheme, a marketplace fee, a moderation policy: each one changes what people gain from cooperating.

Conflux is how I practise thinking about second-order effects before they show up in a metric.

← All work
Personal buildOpen source

DeepJS

A data centre has more jobs than machines. I taught a reinforcement-learning agent to order them better than the standard rules do.

Code
My role
Algorithm, reward design, simulator
Where
IIIT Naya Raipur
When
2023
Built with
TensorFlow, SimPy, Python

01What I changed

  • A steadier learner. I replaced the original policy-gradient method with PPO, which trains more reliably.
  • A two-part reward. The agent is scored on average job completion time and on average slowdown together, so it can't win on one by sacrificing the other.
  • A simulator that takes any job shape. Built on a strategy pattern, it handles jobs with dependencies between tasks and jobs without.

02Result

Time to finish every job, in seconds. Lower is better.

DeepJS with PPO626
First-Fit680
Tetris689

Bars start at zero. The agent finished 8% sooner than First-Fit and 9% sooner than Tetris on the scheduling benchmark.

← All work
ResearchPublished, IEEE Transactions on Education

Learning games

Can a game teach a teenager to think like a programmer? I built four, tested them properly, and published the answer.

Paper on IEEE XploreGoogle Scholar
My role
Game design, build, study design, analysis
Where
IIIT Naya Raipur
When
2024
Built with
Unity, Swift, ARKit
4
games, each mapped to one thinking skill
50
students in classes 9 to 12 in the study
11
citations since publication in IEEE Transactions on Education

01The product

Four games, each designed backwards from the skill it had to build.

Algorithmic, computational and analytical thinking each got their own game. One is an augmented-reality planetarium. Playing well and learning the skill were designed to be the same thing.

02The test

I measured 50 students before and after. Paired t-tests and a one-way ANOVA confirmed the gains were statistically significant.

The framework was published in IEEE Transactions on Education, which accepts 15 to 20% of what it receives.

03The same idea, in a classroom

As lead teaching assistant for 150+ students, I applied the same principle of meeting learners at their level. I grouped students by skill, gave each group its own tasks, and ran a 14-problem coding contest and a Kaggle challenge.

Beginner scores rose from 45% to 70% within the term. Problem-solving speed rose 25% and task completion 20%. I wrote up the experiment here.