Source: https://dhruvipaprunia.com/

Applied AI researcher

![Dhruvi Paprunia at her NMIMS graduation](https://dhruvipaprunia.com/img/27-dhruvi.jpg)

# Hi :) I'm Dhruvi I teach small models to do big things

I explore how AI learns, reasons and acts, and turn that research into systems people can use. At JioHotstar, I build applied AI for real users. Outside work, research is my love.

 [Read my research ↓](https://dhruvipaprunia.com/#research)  [Read a paper breakdown →](https://dhruvipaprunia.com/blog/)  [View CV ↗](https://dhruvipaprunia.com/5-Dhruvi-Paprunia-CV.pdf)  

Selected research · first author [Paper ↗](https://arxiv.org/abs/2509.04518)

**Teaching small models to call tools**

Before training**6.1%**

After RL training**71.1%**

0% 100%

Qwen2.5-3B · 1,000 held-out examples · tool-call accuracy

[CS Best Researcher AwardComputer Science  
NMIMS · 2026 ↗](https://dhruvipaprunia.com/#awards)   [10 %  Dean's Student ListCertificate of Merit  
Top 10% of students ↗](https://dhruvipaprunia.com/#awards)

Say hi 👋

[Email ↗](mailto:dhruvipaprunia1505@gmail.com)  [LinkedIn ↗](https://in.linkedin.com/in/dhruvi-paprunia-0d1505)  [X ↗](https://x.com/Dhruvi1505)  [WhatsApp ↗](https://wa.me/918291206274)

RL for language models*✦*small models, big things*✦*tool use*✦*agents*✦*RAG*✦*4 AI research papers*✦* RL for language models*✦*small models, big things*✦*tool use*✦*agents*✦*RAG*✦*4 AI research papers*✦*

About

## What happens if I try?

My path into research 2022 to 2026 

Underlined links open more ↗

2022 Beginner

I didn't know how to code

2023 Getting started

First Python project   [AI/ML internship at Databyte: multi-camera CCTV tracking](https://dhruvipaprunia.com/#w-cctv)

2024 Discovering research

First research paper   Research head, campus IEEE Robotics & Automation Society   CCTV work became a paper   Began GRPO experiments for small-model tool use

2025 AI researcher arc

[Published our paper on GRPO tool use, taking accuracy from 6.1% to 71.1% ↗](https://arxiv.org/abs/2509.04518)   [Fashion AI: actionable style recommendations ↗](https://ieeexplore.ieee.org/document/11507334/)   [ADTS: crime detection and tracking ↗](https://ieeexplore.ieee.org/document/11118424/)   AI/ML internship at PshPyi Labs   [Built a news recommender on Kubernetes](https://dhruvipaprunia.com/#w-news)   Joined JioHotstar

2026 Applied AI arc

Graduated from NMIMS   [Text-to-SQL chatbot, 96% accurate](https://dhruvipaprunia.com/#w-sql)   Cloud costs: 4 days → 5 min   [HR coach + sales agent](https://dhruvipaprunia.com/#w-hr)   Promoted to Senior Executive, Data Science & Analytics

These days, most of my work comes down to two things:

one **getting small models to do more than people expect**

two **making sure the AI I build still holds up once real users get their hands on it**

[**4 AI research papers** Explore the work ↓](https://dhruvipaprunia.com/#research) 

**6.1% → 71.1%** SLM tool calls with GRPO

**61% vs 43%** Seasonal colour

**0.69 MAE** Cross-domain recommendations

**83 FPS** Real-time tracking

Small models + RL

71.1% from 6.1%

before after GRPO

A strict reward taught a 3B model to call tools, on one T4 GPU.

[Read the paper →](https://arxiv.org/abs/2509.04518)

Text-to-SQL · JioHotstar

12 seasons

Ask a sports question and it writes the SQL, across Pro Kabaddi in five languages and ball-by-ball T20.

[See the project →](https://dhruvipaprunia.com/#w-sql)

RAG + agents

2 agents

An HR coach that answers from handbooks, and a sales agent that picks shows for a brand and writes the deck.

[HR coach →](https://dhruvipaprunia.com/#w-hr) [Sales agent →](https://dhruvipaprunia.com/#w-sales)

Research

## 4 AI research papers

Papers I once barely understood slowly turned into papers I wrote. Here's what each one is about, in plain words.

First author · AIC

**6.1% → 71.1%** TOOL CALLS

### [Advancing SLM Tool-Use Capability using Reinforcement Learning](https://ieeexplore.ieee.org/document/11211997/)

- Cheap small models need exact tool-call formatting.
- GRPO training: any extra text earns zero reward.
- A 3B model reached 71.1% tool-call accuracy on one T4 GPU.

[Read the paper ↗](https://ieeexplore.ieee.org/document/11211997/) [arXiv ↗](https://arxiv.org/abs/2509.04518)

First author · RCSM

**61% vs 43%** SEASONAL COLOUR

### [Towards Actionable Fashion AI: A Holistic, End-to-End Style Recommendation Assistant](https://ieeexplore.ieee.org/document/11507334/)

- Seasonal colours and body shape become 20-25 clothing picks.
- Fine-tuned DINOv2: 61% seasonal-colour accuracy versus Gemini 2.5 Flash at 43%.
- Designed to run on a phone.

[Read the paper ↗](https://ieeexplore.ieee.org/document/11507334/)

Co-first author (equal contribution) · SSITCON

**0.69 MAE** RECOMMENDATION ERROR

### [LightFusionRec: Lightweight Transformers-Based Cross-Domain Recommendation Model](https://ieeexplore.ieee.org/document/10796005/)

- Recommend books from movies a user liked, even with little history.
- Combine DistilBERT text features with FastText genre features.
- MAE (mean absolute error): 0.69 at top 20%, beating text-only, genre-only and TF-IDF baselines.

[Read the paper ↗](https://ieeexplore.ieee.org/document/10796005/) [arXiv ↗](https://arxiv.org/abs/2410.15656)

Second author · ICOCT

**83 FPS** REAL-TIME TRACKING

### [Advanced AI-Based Detection and Tracking System (ADTS) for Crime Prevention and Identification in Real-Time Surveillance](https://ieeexplore.ieee.org/document/11118424/)

- Detect people, recognise faces and track them across CCTV cameras.
- YOLOv8, MTCNN and FaceNet: 83 frames per second.
- IDF1 (identity-tracking score): 89.8%.

[Read the paper ↗](https://ieeexplore.ieee.org/document/11118424/)

Awards & roles

## Beyond the papers

College was never just about academics for me.

- **Research Head** , IEEE Robotics & Automation Society, NMIMS
- **Student Coordinator** , Student Research Group, NMIMS
- **Sub-Head** , Google Developer Student Clubs, NMIMS (400+ students)

Work

## Things I've built

A mix of things running in production and experiments I just wanted answers to.

JioHotstar

### Text-to-SQL sports chatbots

96% accurate

- Started from zero, with a lot of research before building
- 12 seasons of Pro Kabaddi in English, Hindi, Hinglish, Gujarati or Marathi, plus ball-by-ball T20
- 96% right on our 150-question Kabaddi test

JioHotstar

### Cloud cost dashboard

4 days → 5 min

- ~62 projects across GCP and AWS in one view
- About 4 days of manual work down to about 5 minutes
- Senior management uses it now

JioHotstar

### HR AI Coach

Hybrid RAG

- Answers questions from company handbooks
- Keeps HR record facts separate from anything the AI writes, so the two never mix

JioHotstar

### Sales Agent

Top 5 shows

- Takes a brand's brief and picks the 5 best-fit shows
- Writes the pitch deck around them
- The sales team uses it now

Project

### Promo effectiveness dashboard

6 sources

- Shows how a promo is actually landing
- Pulls signals from YouTube, Instagram, X, Reddit, Google Trends and Hotstar
- Reads the comments for sentiment

Project

### Episode condensation

43% shorter

- Works from the dubbing script
- Trims recaps, ad-break gaps and pauses
- One episode went from 24:33 to 13:56

JioHotstar

### Audio separation pipeline

6 stems

- Benchmarked SAM-Audio against HTDemucs on a real production file; BS-RoFormer won out
- Splits TV audio into 6 stems, so crowd noise, applause and camera clicks come out while the music stays
- Ran on masters from 4 channels

JioHotstar

### Music catalog downloads

~10k audio files

- Crawls a music library's full genre tree and saves each track as WAV to cloud storage
- Playwright automation, with ProcessPool running workers in parallel and systemd keeping long bulk downloads going
- A download lock so workers never clash, and every file checked against its track title and catalog ID

JioHotstar

### Music copyright detection

180+ songs

- Tested ACRCloud, AudioShake and Gemini against cue sheets I checked by hand
- AudioShake did best, 80.8-90.1% on music-only audio
- Moved timestamps out of the LLM and into Python, which took Gemini to 64%

JioHotstar

### HR onboarding videos

5 models tested

- Tried 5 lip-sync models, picked LatentSync
- With ElevenLabs voices and n8n, makes personalized onboarding videos in batches

PshPyi Labs · internship

### Personal branding agent

n8n agents

- Co-built an LLM platform on n8n that writes, schedules and publishes social posts
- I designed the agent workflow nodes that make each post personal

Project

### Personalized news recommender

Java + K8s

- Java and Spring Boot microservices with service discovery and an API gateway
- Runs on Docker and Kubernetes
- Recommendations from TF-IDF and each user's preferences

Databyte · internship

### Multi-camera CCTV tracking

→ research paper

- Real-time face tracking across multiple CCTV feeds
- YOLOv8, MTCNN and FaceNet
- Became the ADTS paper

Side projects

## Built for fun

- **Hacker News → tweets.** Every day it reads the top 15 Hacker News stories, sums them up and drafts 10 tweets from them. I built it with CrewAI and Streamlit. [GitHub ↗](https://github.com/dap15/Hacker-News-Twitter-Text-Generator)
- **Salary negotiation bot.** It predicts your salary with ML, and then you haggle with an LLM bot over it. After that, a second bot negotiates the same profile with a boss bot, so you can see who got the better deal. [GitHub ↗](https://github.com/dap15/Salary-Negotiation)

Writing

## Explained simply

[Paper deep dive

### Your model found the right answer. It might never learn to find it again.

What GRPO, a reward-based training method, changes across several tries.

15 min read + optional walkthroughs Reward training

![Rare correct proof](https://dhruvipaprunia.com/blog/a-rare-correct-answer-can-still-get-lost-in-training/cover.png)](https://dhruvipaprunia.com/blog/a-rare-correct-answer-can-still-get-lost-in-training/) [Lab deep dive · Ai2

### The model weights are open. Can we see how the model learned?

Why open model training matters, and how to test the steps that shape an assistant.

8 min read Open models

![Open training notebook](https://dhruvipaprunia.com/blog/the-model-weights-are-open-can-we-see-how-the-model-learned/cover.png)](https://dhruvipaprunia.com/blog/the-model-weights-are-open-can-we-see-how-the-model-learned/) [Paper deep dive

### A million tokens can still lose the right document

Why seeing every document is not the same as finding the right one.

10 min read Retrieval

![Wall of documents](https://dhruvipaprunia.com/blog/a-million-tokens-can-still-lose-the-right-document/cover.png)](https://dhruvipaprunia.com/blog/a-million-tokens-can-still-lose-the-right-document/)

[All writing →](https://dhruvipaprunia.com/blog/)

Say hi

## There's always another question hiding behind the answer. Send me yours.

[dhruvipaprunia1505@gmail.com](mailto:dhruvipaprunia1505@gmail.com) [LinkedIn](https://in.linkedin.com/in/dhruvi-paprunia-0d1505) [X](https://x.com/Dhruvi1505) [GitHub](https://github.com/dap15) [Download CV (PDF)](https://dhruvipaprunia.com/5-Dhruvi-Paprunia-CV.pdf)
