Portfolio

DHRUV

move your cursor  ·  click to scatter
scroll
About

I teach small models what the giants know — then I put them to work.

I'm Dhruv, a software developer working where modern AI meets the enterprise back office. My core craft is knowledge distillation: transferring the capability of massive models into compact ones that run fast, private, and anywhere.

At Metavision Technology I ship that thinking into real businesses, building AI and agentic workflows on top of Tally and ERP systems used by logistics, biotech and finance companies. Legacy systems hold the knowledge; modern AI passes it on. The fresco above is the job description — one hand passing the spark to another.

Dhruv Shah
Case study

The Distillation

My flagship work: compressing a fine-tuned GPT-2 Large into DistilGPT-2 without losing its mind. A pipeline of semantic filtering, dual-LLM deduplication, multi-model augmentation and dual-loss training: 774 million parameters taught into 82 million. Keep scrolling and watch it happen.

Teacher  ·  GPT-2 Large  ·  774M parameters Student  ·  DistilGPT-2  ·  82M parameters
Capability Preserved
Complexity Removed
Results

Smaller. Faster. Still smart.

89%
fewer parameters
up to 8×
faster inference
4.7×
smaller on disk
github.com/Dhruvshah0506/Knowledge_Distillation
Now building

ARKA‑BI In development

Metavision Technology  ·  Agentic business intelligence

One centralized place for a business to understand itself. ARKA‑BI plugs directly into Tally and puts agents to work on the pain points every finance team knows too well: reporting by hand, answers buried in ledgers, problems discovered a month late.

Ask, don't queryPlain-language questions answered from live Tally and ERP data.
Reports that build themselvesMIS and business reports assembled by agents, not by hand in Excel.
Signals, not spreadsheetsAnomalies in receivables, cash flow and trends surfaced before they hurt.
Agentic pipelinesMulti-step workflows: fetch from Tally → transform → analyze → deliver.
Experience

Metavision Technology  ·  Software Developer

2026 – Present  ·  Mumbai

Metavision blends legacy enterprise systems with modern AI: Tally, SAP B1 and MIS platforms on one side, agentic workflows on the other. I develop and maintain the systems that pull data out of Tally and put it to work: custom client systems, MIS analytics, internal data pipelines, and ARKA‑BI, our agentic reporting platform. The throughline from my research is the same: take heavyweight capability and make it run where the business actually lives.

Experience

Tally Solutions  ·  AI/ML Intern

May 2025 – August 2025

Engineered a teacher-student knowledge distillation pipeline, compressing a massive model's knowledge into a domain-specific Small Language Model (SLM) for accounting and finance. Reduced inference time by 40% while retaining 76% of the teacher's accuracy. Authored automated NLP preprocessing pipelines (Pandas, NLTK) for 3,000+ unstructured financial texts, boosting training efficiency.

Research

Shortcut Learning in Deep Neural Networks

Published Preprint  ·  2026

First-authored a paper on detecting, characterizing, and mitigating shortcut learning in automated Knee Osteoarthritis (OA) grading. Engineered a three-phase experimental framework utilizing a disentangled ResNet-18 architecture extended with a Gradient Reversal Layer (GRL). Successfully mitigated domain-sensitivity and shifted model focus to true diagnostic anatomical features.

Read the paper

Stack

Tools of the Trade

Languages: Python, SQL, JavaScript, C#, Java, C++, PHP
ML Frameworks: PyTorch, Hugging Face Transformers, Scikit-learn, Pandas, NumPy
Core Competencies: Small Language Models (SLMs), NLP, Knowledge Distillation, Agentic Workflows, Adversarial Disentanglement.
Engineering: MERN Stack, .NET, Tally / ERP Integration, Data Pipelines, Model Optimization.

Craft

What I do

Knowledge distillationTeacher–student pipelines that keep the capability and drop the parameters.
SLM trainingPre-training and fine-tuning small language models, end to end.
EvaluationBenchmarks that prove the student learned, not memorized.
Agentic systemsMulti-step agents that fetch, reason over and deliver enterprise data.
DeploymentModels sized for real hardware and real latency budgets.
Contact

Let's create something together.

Open to interesting problems in applied ML and enterprise software. The fastest ways to reach me:

after Michelangelo · Sistine Chapel · c.1512