Sanyam Lamba — Building, Experimenting, Learning.

02 / Journey

A Story in Progress

Not a résumé. A narrative of curiosity.

2022

The Beginning

Started Computer Science at LNMIIT. First real exposure to algorithms, data structures, and the quiet realization that software is organized thinking at scale.

LNMIIT · Jaipur

2024

First Deep Dive into AI

Built an Accident Detection system using Deep Learning — VGG16, transfer learning, custom loss functions. First encounter with Alpha Focal Loss and Xtreme Margin Loss. Realized quickly that AI design is more craft than formula.

Deep Learning · Computer Vision

2025

Research Meets Privacy

Worked on Federated Client Selective Unlearning for Medical Imaging. Deep experimentation with FFT, DFT, DWT, and FWT transforms. The question that kept me up at night: how exactly do you teach a model to forget?

Federated Learning · Privacy-Preserving AI

2026

The Expanding Horizon

Exploring CTGANs, Synthetic Data Generation, LLM Workflows, and AI Automation. Each thread leads to three more questions. Currently building more than I’m reading — and that feels right.

Generative AI · LLMs · Automation

03 / Projects

Featured Work

A case study for each project — problem, approach, experiment, lesson.

01 2026

Synthetic Tabular Data Generation using CTGAN

How do you generate data that doesn’t exist yet — but statistically could? A GAN-based approach to synthesizing tabular datasets with class imbalance handling.

Generative AI CTGAN Data Synthesis Statistical Validation

The Problem

Real-world tabular datasets are often imbalanced, privacy-sensitive, or simply too small to train robust models. The challenge was generating synthetic data that preserves statistical properties while handling severe class imbalance.

The Approach

Used Conditional Tabular GAN (CTGAN) — a GAN architecture designed for mixed-type tabular data. CTGAN uses mode-specific normalization for continuous columns and conditional generation to address imbalanced categorical distributions.

Experimentation

Ran statistical validation across column-wise distributions, correlation matrices, KS tests, and downstream ML performance comparison between real and synthetic+real training sets. The core question: does synthetic data help or just add noise?

What I Learned

Synthetic data is about augmentation, not fabrication. The hardest part wasn’t generating data — it was knowing when it was good enough. Statistical similarity doesn’t guarantee ML utility. That gap is the real research problem.

02 2025

Federated Client Selective Unlearning

Can a model learn to forget? A research exploration into privacy-preserving AI, selective forgetting, and frequency-domain weight analysis in federated settings.

Federated Learning Machine Unlearning Medical Imaging FFT · DWT

The Problem

In federated learning, clients train a shared model without sharing raw data. But what happens when a client wants to be “forgotten” — when GDPR demands data deletion? Full retraining is computationally prohibitive at scale.

The Approach

Explored selective unlearning strategies that let the global model forget a specific client’s contribution without full retraining. Applied in medical imaging, where privacy is a hard constraint, not a preference.

Experimentation

Experimented with FFT, DFT, DWT, and FWT transforms for frequency-domain analysis of model weights. Investigated whether weight-space transformations reveal patterns that make selective forgetting more precise and measurable.

What I Learned

Machine unlearning is deceptively hard to verify. How do you prove a model has forgotten? The evaluation metrics are themselves a research problem. Privacy in AI is not a feature to add — it’s an architectural concern from day one.

03 2024

Accident Detection using Deep Learning

Transfer learning meets real-world safety. A computer vision system that detects road accidents from video frames, with custom loss functions for severe class imbalance.

VGG16 Transfer Learning Alpha Focal Loss Xtreme Margin Loss

The Problem

Accident detection from video feeds is a critical safety application — but accidents are rare events. Highly imbalanced datasets mean standard training fails. Standard cross-entropy loss treats every frame as equally important.

The Approach

Used VGG16 as backbone with transfer learning from ImageNet. The innovation was in the loss functions: implemented and compared Alpha Focal Loss and Xtreme Margin Loss for forcing the model to attend to rare accident events.

Experimentation

Systematic comparison of loss functions, fine-tuning strategies, and data augmentation. Each experiment revealed new intuitions about how the model weighted rare events. Benchmarked across precision, recall, F1, and AUC.

What I Learned

In safety-critical systems, missing a positive is worse than a false alarm — recall over precision. Model design is inseparable from domain knowledge. You can’t optimize a loss function without understanding what failing costs.

04 / Learning

What I’m Exploring

A notebook of ongoing curiosities. Not mastery — investigation.

♦

Generative AI

From VAEs to diffusion models. The art of controlled randomness and what “generation” really means.

Active

◦

LLM Agents

Building systems that reason, plan, and act. What does “autonomous” actually mean at a systems level?

Exploring

◆

LangChain

Orchestrating language models. Chains, agents, memory, retrieval. Plumbing for intelligent systems.

Active

△

Automation Workflows

If it’s repeatable, it should be automated. AI-powered process design that actually saves time.

Building

◯

Synthetic Data

Creating data that doesn’t exist yet — GAN-based generation for tabular and image domains.

Deep Dive

□

Machine Unlearning

The art of selective forgetting. Privacy by design, not as a patch applied after the fact.

Researching

✦

Deep Learning

Architectures, loss functions, regularization. The fundamentals keep revealing new depth the more you use them.

Ongoing

◉

Prompt Engineering

Communicating intent to language models precisely. More craft than science — context is everything.

Practicing

05 / Beyond Code

The Human Side

What I do when I’m not debugging models or reading papers.

Leadership & Events

TEDxLNMIIT

Event Management Lead

Coordinated one of the most complex student-led events at LNMIIT. Managed cross-functional teams, drove sponsorship conversations, and ensured that the logistics of bringing big ideas to a stage actually worked — invisibly.

What I carried forward

Leadership is mostly about clearing obstacles for your team, not directing from above
Sponsorship conversations are a masterclass in articulating value clearly
The best events feel effortless precisely because of invisible preparation

Community & Teaching

Sankalp

Core Member

Part of a student initiative to help support staff at college — mess workers, housekeeping — prepare for competitive exams. Teaching someone who genuinely wants to learn is one of the most rewarding things I’ve done at university.

What I carried forward

Explaining something simply means you understand it deeply — no shortcuts
Motivation matters far more than method in any teaching relationship
Community work reframes what “success” looks like in meaningful ways

06 / Tech Stack

The Workshop

Tools I reach for regularly. Not trophies — instruments.

Languages

Python

SQL

Frameworks

TensorFlow

PyTorch

LangChain

Libraries

Pandas

NumPy

OpenCV

CTGAN

Tools

Git

GitHub

Jupyter

Concepts

Transfer Learning

Federated Learning

Machine Unlearning

Synthetic Data

07 / Thoughts

Writing (Soon)

Ideas forming. Posts pending. Watch this space.

Essay

Why Synthetic Data Matters More Than You Think

On data scarcity, privacy constraints, and the gap between what we have and what we need to train models that actually generalize.

Draft in progress

Technical

Understanding Machine Unlearning

The right to be forgotten, applied to neural networks. Why selective deletion is an unsolved research problem — not an engineering task.

Notes phase

Reflection

Lessons from Building Deep Learning Projects

What papers don’t mention: failed experiments, strange loss curves, and the intuitions that only come from actually building.

Outline ready

Research

What Research Papers Don’t Tell You

The gap between a published result and a working implementation. On reading papers vs. reproducing them.

Thinking phase

08 / Contact

Let’s Talk

I’m always interested in conversations about AI, research, experiments, and interesting ideas. If you’re working on something in the space of generative AI, federated learning, or just want to think out loud — reach out.

Email sanyamlamba2002@gmail.com LinkedIn linkedin.com/in/sanyam-lamba GitHub github.com/sanyam-lamba