Divyam Sharma
AI engineer who takes LLM products from prototype to launch. I built Savra's AI exam-paper system as its only engineer, and did product work at Paytm.
33 subjectsof CBSE exam papers Savra's AI now writes, for grades 6–12 in English and Hindi
10× cheaperto make each paper, after a blind test of four designs at Savra
23% → 0%shallow, recall-only questions in Savra's papers
~70%less partner onboarding effort with the portal I built at Paytm
15+ chartsin the Superset dashboard Paytm's leadership uses to pick features
work
Forward Deployed AI Intern at Savra
AI startup in educationAug – Nov 2026shipped
- Took Savra's AI exam-paper product from a prototype that had never reached a teacher to a launch covering 33 subjects across grades 6–12, in English and Hindi, including 1,300+ diagram questions.
- Tested four approaches blind before committing. The result overturned the design I inherited and cancelled a sprint that wouldn't have helped; papers became 10× cheaper to make, and shallow questions fell from 23% to 0%.
- Built the checks every paper goes through: an AI grader tested against real CBSE board questions, and an answer-key checker that re-solves each question and fixes what's wrong.
- Turned 476 textbook PDFs into 26,000+ clean, searchable passages, found and fixed a hidden defect that was corrupting 12.8% of them, and cached each chapter's key facts so no paper starts from scratch.
- Rolled it out to selected teachers first, over 7 staged releases, with an automatic switch back to the old system if anything failed. Alongside, shipped ~100 changes to the live app, from a credits system to offers that won back lapsed teachers.
PythonFastAPIGemini APIQdrantRedisPostgreSQLNode/ExpressDocker
Product Intern at Paytm
Gift Vouchers and Paytm Loves EMI · NoidaJun – Jul 2026
- Replaced manual document-sharing with a Gift Voucher integration portal for resellers and brands, cutting partner onboarding effort by ~70%. It became the standard process for every new integration.
- Built the Paytm Loves EMI dashboard from scratch in Apache Superset, with 50+ SQL queries and 15+ charts over ~6 lakh monthly events. Leadership uses it to decide which features come next.
- Found why 10.5% of searches (out of 5 lakh a month) came back empty, sized the impact, and worked with the Catalog API team to ship a fuzzy-matching fix.
- Defined ~40 tracking events for the EMI app, which later funnel analysis was built on, and turned customer research and dashboard insights into 12 product ideas.
- Looked after 3 major brand partners (Rs 10 Cr+ in potential business), turning what they needed into working integrations.
SQLApache SupersetTrinoFunnel analysisEvent trackingJira
Product Intern at Stotio
startupAug – Oct 2025
- Went through the product's user flows to find where people got stuck.
- Turned what I found into recommendations and clear requirements for new features.
- Helped run outreach across channels to grow the brand.
User-flow analysisUX researchProduct requirementsOutreach
Software Development Intern at VVDN Technologies
Mobile and Cloud Apps · NoidaJun – Aug 2025
- Owned a role-based task-management system with Admin, Manager and Employee panels, taking 10+ features from spec to production.
- Shipped 10+ production REST APIs in Django REST Framework, with SimpleJWT login and custom middleware that checks who is allowed to do what.
- Designed the database so new features could be added without breaking existing data.
PythonDjangoDjango REST FrameworkSimpleJWT
projects
Relay: an enterprise AI gateway
personal projectJul 2026new
- Built a self-hosted gateway that gives a company one OpenAI-compatible API for every AI provider, so its apps switch over by changing a single URL.
- Every request is checked on the way in and on the way out: 31 rules against jailbreaks and prompt injection plus a check that catches reworded attacks, 8 rules that strip personal data from answers, and login and rate limits.
- Built a router that sends simple prompts to cheaper models and reuses answers to repeated questions, cutting cost by 43.9% on a 32-prompt benchmark.
- Shipped it with 149 automated tests, including 61 regression cases, and a dashboard that shows each request's cost, speed and safety decisions as they happen.
PythonFastAPILiteLLMChromaDBRedisPostgreSQLReactDocker
CiteScope: a citation finder that knows when it can't answer
information retrieval projectOct 2026new
- Built a citation finder over 26,601 research papers: give it an abstract and it lists the papers that abstract should be citing.
- Wrote the search engine from scratch, without a search library, then tested it against three other approaches. The best mix found the right papers 21% more often than the basic version, across 400 test abstracts.
- When the right papers aren't in the collection, it says so instead of guessing. It's right 95.4% of the time it declines, and it turned away 148 of 150 abstracts from unrelated fields.
- Every 'I can't answer this' comes with a reason, naming the terms the collection doesn't cover, so a person can check it.
PythonNumPyNLTKsentence-transformersOpenAlex APIStreamlit
WHAT IF? An engine that keeps an AI's facts and guesses apart
LLM trust and verification · personal projectMay 2026
- Language models blend facts and guesses into one confident answer. WHAT IF? keeps them apart, so every claim can be checked against where it came from.
- Built a five-agent LangGraph pipeline (understand, retrieve, ground, reason, score) that tags every claim with the passage it came from and a confidence score.
- Confidence comes from how much evidence backs each step, not from asking the model how sure it feels. Citations it makes up get flagged instead of trusted.
- Searches 1,466 Wikipedia passages with local embeddings, pulls in live Wikipedia when nothing matches well enough, and stops after 4 reasoning steps so it can't drift into fiction.
PythonLangGraphChromaDBPydanticsentence-transformersCerebrasStreamlit
LexiGrade: a reading-level classifier
NLP project with a web appApr 2026
- Built a classifier that grades any passage from Beginner to Expert, with a small Flask web app to try it on your own text.
- Paired a LightGBM model, trained on 1,000 labelled passages, with six classic readability formulas, so it looks at what the text says and not just how long the words and sentences are.
- Fixed a blind spot those formulas share: short but dense sentences, which they wrongly rate as easy.
PythonLightGBMall-MiniLM-L6-v2Scikit-learnFlask
Subliminal learning in LLMs
research replicationJan 2026
- Replicated 'subliminal learning' (arXiv 2507.14805), the finding that a model can pick up another model's traits from data that has nothing to do with them.
- Told a teacher model it loves owls, had it write 10,000 sequences of plain numbers, and fine-tuned a Qwen2.5-3B student on those numbers alone.
- Fully fine-tuned, the student picked the owl in 13.4% of 500 answers. A cheaper LoRA-trained student showed no change, so how the student is trained decides whether the trait gets through.
- Ran the same experiment on Gemma-2B, training only 0.07% of its weights.
PythonPyTorchHugging Face TransformersPEFT / LoRAQwen2.5-3BGemma-2B
Smart Price: pricing Amazon products from their listings
Amazon ML Challenge 2025Oct 2025
- Fine-tuned NeoBERT to predict prices for 75,000 Amazon products from messy, inconsistent listing text.
- Trained it on log-scaled prices with a loss matched to how the challenge was scored, so being 10% off counts the same on a cheap item as on an expensive one.
- Built a second version that also looks at the product photo, combining what the image and the text say.
- Tracked every training run in MLflow and wrote a script that combines several models' predictions into one.
PythonPyTorchTransformersNeoBERTCLIPMLflow
EcoBin: real-time waste sorting on a Raspberry Pi
edge ML, team projectAug 2025
- Trained a MobileNetV2 model on 60,000+ photos of waste to sort it into three classes, with 97% accuracy.
- Shrank it with TensorFlow Lite so it runs on a Raspberry Pi, taking under 100 ms per item without sending anything to the cloud.
PythonTensorFlowMobileNetV2TensorFlow LiteRaspberry Pi
GroomMe: an AI stylist for clothes you already own
personal projectAug 2025
- Built the backend for an AI stylist: tell it the occasion and it puts together an outfit from your own wardrobe, not a shopping list.
- Gemini first plans the ideal outfit for the occasion, then rebuilds it using only the user's clothes, ranked by style, occasion and season.
- Built the Django REST API with email OTP sign-up and JWT login, and recorded each piece of clothing's colour, material, season and style, so suggestions rest on real details.
PythonDjangoDjango REST FrameworkGemini APIPostgreSQLSimpleJWT
education and community
B.Tech in Computer Science at Shiv Nadar University
major · AI/ML specialisation · CGPA 8.562023 – 2027
B.Tech in Computer Science at Shiv Nadar University, specialising in AI and ML. Graduating in May 2027.
Minor in Economics
Shiv Nadar University · GPA 8.8
A minor alongside CS, covering microeconomics, macroeconomics, game theory and Indian financial markets.
St. Dominic's Senior Secondary School
Mathura Cantt · CBSEClass of 2023
Where I went to school, a convent school in Mathura, through Class XII.
Teaching Assistant for DSA and Operating Systems
370+ students2025 – 2026
Teaching assistant for Data Structures and Algorithms (Aug – Dec 2025) and Operating Systems (Jan – Apr 2026), for 370+ students in all.
Student Ambassador
2026
Chosen to represent the university at the Industry Partners Summit 2026, where I moderated an AI roundtable with CTOs and senior partners.
Marketing Lead at E-Cell
Led a 40-member team running E-Cell's social media and outreach, growing our reach from 5K to 20K+ impressions.
Curation Lead for XCELERATE 3.0
Headed curation for XCELERATE 3.0, a national ideathon with 250+ submissions. Led a team of 30 and designed how business plans were judged to pick the top 50 finalists.
off duty
University volleyball team
Played for the university volleyball team.
Singing
Usually humming something. Sometimes on a mic.
Dancing
Give me a good song and a dance floor.
Cooking
Cooking, and eating good food.
Movies and Friends reruns
Movies, and Friends on repeat.
Poker nights
Poker nights with friends: odds, reads and the occasional bluff.
Badminton
Always up for a game.
Cricket
Bat, ball, and an argument about the rules.
skills
ai and llm systemsRAGLangGraphLangChainLLM evaluationInformation retrievalEmbeddingsQdrantChromaDBGemini API
machine learningPyTorchTransformersFine-tuningLoRA / PEFTComputer visionCNNsTensorFlowTFLiteScikit-learnLightGBM
backendFastAPIDjangoDjango REST FrameworkNode/ExpressJWTRedisDocker
dataPostgreSQLSQLPandasNumPySQLiteMySQLSupabase
product and analyticsProduct managementFunnel analysisEvent trackingSuperset
languages and frontendPythonTypeScriptJavaScriptC++CReactHTMLCSS