Siddharth Umathe | Applied AI Engineer

Siddharth
Umathe

I build AI systems that research, reason, retrieve, evaluate and ship.

Applied AI across agentic systems, RAG, speech, multimodal AI and computer vision. I focus on turning ambiguous problems into working systems with deliberate architecture, measurable evaluation and product level engineering.

01 / FLAGSHIP SYSTEM

Best Software Engineering Project GenAI · RAG · Product

AI Powered Software Engineering System

AI Engineer · Team of 7 · IIT Madras

Designed and integrated a broad, multi module AI layer for an academic software platform. The system combined retrieval grounded assistance, lecture and weekly summarisation, notes and assessment generation, debugging support and topic recommendations inside a single product workflow.

Python Flask SQLAlchemy Gemini API LangChain ChromaDB RAG Structured Outputs

Problem

Students repeatedly switch between lecture content, notes, revision, debugging and assessment preparation. The goal was to make those workflows context aware and available inside one academic system rather than as isolated AI prompts.

My role

Worked as AI Engineer in a seven member team, leading the AI module design and integration across GenAI pipelines, retrieval architecture, prompt design and backend orchestration.

System modules

Contextual AI assistant
Lecture and video summarisation
Week level summarisation
Topic notes generation
Practice question and mock quiz generation
Coding assistant for error explanation
Topic recommendation support

Engineering depth

Worked across transcript processing, content chunking, retrieval using ChromaDB, LLM prompting, structured outputs, backend API integration and response reliability handling. The emphasis was system integration rather than isolated generation.

Outcome

The project was recognised as the Best Software Engineering Project, providing external validation for the overall product and engineering execution.

What changed in my thinking

This project moved my focus from using an LLM to engineering complete AI workflows: context management, modularity, data flow, failure handling and user facing product integration.

Speech AI

Automatic Speech Recognition Systems

Self supervised speech · ASR pipelines

Built end to end speech recognition pipelines around wav2vec2.0, HuBERT and Whisper, with attention to preprocessing, variable length audio, decoding, GPU aware training and transcription quality.

PyTorch wav2vec2.0 HuBERT Whisper Hugging Face CTC

Problem

Speech systems must handle variable length audio, sampling rate differences, noisy inputs and alignment between acoustic representations and text tokens.

Approach

Worked on waveform normalisation, tokenisation, batching, dynamic padding, alignment strategies, CTC based decoding and encoder driven ASR workflows.

Engineering depth

Designed preprocessing and model pipelines with memory conscious batching, Hugging Face Transformers and PyTorch experimentation for stable training and inference.

Learning

Built practical understanding of speech representation learning, sequence modelling, CTC alignment, inference behaviour and ASR pipeline design.

LLM Adaptation

Google Gemma Fine Tuning with LoRA and PEFT

Parameter efficient fine tuning · NLP

Adapted Google Gemma models for domain specific NLP tasks using LoRA and PEFT, covering data preparation, prompt formatting, tokenisation, training configuration and inference evaluation.

PyTorch Hugging Face PEFT LoRA Gemma

Problem

Full model fine tuning is costly when a model only needs targeted behavioural or domain adaptation. The project explored parameter efficient adaptation as a practical alternative.

Approach

Built the training workflow across dataset preprocessing, prompt formatting, tokenisation, batching, LoRA configuration, training settings and inference testing.

Engineering depth

Worked with Transformers, PEFT, tokenizer pipelines, instruction tuning workflows and GPU backed training environments.

Learning

Developed practical understanding of parameter efficient adaptation, prompt structure, transformer behaviour and inference robustness.

Computer Vision

4× Image Super Resolution

Competition project · Image restoration

Developed CNN based super resolution pipelines for reconstructing higher quality images from low resolution inputs, using residual learning and perceptual evaluation.

PyTorch CNN Residual Learning PSNR SSIM

Problem

Super resolution requires recovering fine spatial detail while avoiding unnatural sharpening and visual artefacts.

Approach

Built 4× image super resolution workflows around CNN architectures, residual learning, feature extraction and perceptual loss concepts.

Engineering depth

Worked across image resizing, normalisation, patch based training, augmentation, GPU aware workflows and evaluation using PSNR and SSIM.

Learning

Built practical intuition for restoration architectures, perceptual learning and visual model evaluation.

Generative Vision

GAN Style Generative Image Modeling

Adversarial training · Generative modeling

Developed and trained GAN style architectures for image generation, focusing on dataset preparation, generator discriminator balance, latent behaviour and training stability.

PyTorch GANs Adversarial Training Latent Space FID

Problem

Adversarial image modelling must balance realism, diversity and training stability while avoiding mode collapse.

Approach

Built data pipelines, augmentation, generator discriminator training loops and GPU backed experimentation around GAN style architectures.

Engineering depth

Worked with encoded image shards, normalisation, batching, generator capacity, discriminator balance, regularisation and FID style evaluation concepts.

Learning

Developed hands on understanding of adversarial optimisation, latent representations and generative model stability.

Applied Data

Business Data Management · Native Chefs

Real world B2C data · Business analytics

Analysed operational data from a home cooked food delivery business to surface unpaid orders, dish performance, customer behaviour and revenue related insights.

Python Pandas Matplotlib EDA Business Analytics

Problem

The business needed clearer visibility into unpaid orders, dish performance, customer ordering patterns and potential revenue leakage.

Approach

Performed cleaning, preprocessing, descriptive analysis, pivot based summaries, visualisation and interpretation against business questions.

Focus areas

Revenue leakage through unpaid orders
Dish level performance
Customer ordering behaviour
Revenue trends
Operational recommendations

Learning

Strengthened the ability to translate raw data into decision support insight rather than treating analysis as an endpoint.

Applied AI is an engineering problem, not a demo problem.

I am an Applied AI Engineer trained at IIT Madras, with focused academic work in Generative AI and Multimodal AI. My work spans agentic and retrieval augmented systems, speech AI, NLP fine tuning, computer vision and applied data science.

I am most interested in problems where research and engineering meet: deciding how a system should retrieve context, structure reasoning, use tools, validate outputs, handle failure and become reliable enough to sit inside a real product.

I also bring commercial operating context from building and marketing digital products. Working across positioning, creative, acquisition and conversion has made me more demanding about what an AI system is actually for, who it serves and how its value becomes visible.

I learn by building, measuring and revising. The strongest projects are not the ones with the longest feature lists. They are the ones where architecture, evaluation and user value reinforce each other.

The same systems lens extends into independent strategic foresight. I study how capability, dependency, replaceability, trajectory, constraints and second order effects shape geopolitics, defence, AI, robotics and long range technological power.

Start from the problem

Model choice comes after understanding what must be solved, verified and delivered.

Design for failure

AI systems need grounding, validation and observable failure modes, not just good happy path demos.

Evaluate the system

Accuracy, consistency, latency, cost and user usefulness all matter depending on the product.

Ship the whole workflow

APIs, retrieval, data flow, orchestration and product integration are part of the AI system.

Built around systems, not keyword lists.

Agentic / LLM Systems

Large Language Models RAG AI Agents Prompt Engineering Structured Outputs LLM Evaluation Gemini API LangChain ChromaDB

Speech / Multimodal

wav2vec2.0 HuBERT Whisper ASR Pipelines CTC Decoding Vision Language Models Multimodal AI

Deep Learning / Vision

PyTorch Deep Learning Neural Networks Computer Vision Representation Learning Generative Modeling RL Foundations

Engineering

Python Flask FastAPI SQLAlchemy REST APIs Docker Git Linux SQL Pandas NumPy

I understand what happens before and after the model.

Applied AI is my technical core. But I have also built and operated the commercial side of digital products across research, positioning, landing pages, performance creative, paid acquisition, tracking and iteration. That changes how I engineer. I think about the user, the feedback loop and the business outcome, not only the model.

Technical depth with a strong applied AI bias.

Degree

B.S. in Data Science and Applications

Indian Institute of Technology Madras

Graduation: 2026

Minor

Generative AI

IIT Madras

Large Language Models Deep Learning Practice Mathematical Foundations

Minor

Multimodal AI

IIT Madras

Large Language Models Speech Technology Deep Learning for Computer Vision

Advanced coursework

Selected technical training

Reinforcement Learning MLOps Data Visualization Design

Applied / business

Commercial context

Market Research Digital Marketing AI in Business

B.Sc. Computer Science · RTMNUFirst Class

Class XII · Maharashtra State Board84%

Class X · Maharashtra State Board91%

Signals that matter beyond self description.

Award

Best Software Engineering Project

Recognition for the IIT Madras software engineering project in which I worked as the team AI Engineer across a broad multi module GenAI system.

Project ownership

AI Engineer · Team of 7

Led AI module design and integration across retrieval, prompting, context handling and backend orchestration inside the team product.

Specialisation

Minor in Generative AI

Focused academic training in LLMs, deep learning practice and mathematical foundations of generative systems.

Specialisation

Minor in Multimodal AI

Focused coursework spanning LLMs, speech technology and deep learning for computer vision.

Operator proof

Owned Product Growth Systems

Personally operated the commercial chain across research, landing pages, performance creative, paid media, tracking and iteration, creating direct feedback between product decisions and market response.

Problems I want to work on.

I am especially drawn to AI systems where reasoning quality, multimodal understanding, human consequences and engineering reliability matter more than producing a flashy demo.

AI Systems

End to end intelligent pipelines and modular AI architecture.

Healthcare AI

Decision support, clinical intelligence and medical NLP.

Multimodal AI

Vision, language and speech across shared reasoning workflows.

Speech AI

ASR, speech representation learning and spoken interfaces.

Agentic AI and RAG

Tool use, retrieval, structured reasoning and grounded generation.

Reinforcement Learning

Sequential decision making and policy optimisation.

Computer Vision

Image understanding, restoration and generative vision.

High Stakes AI / Strategic Foresight

Reliable AI, decision support, national power and long range systems thinking.

Building something difficult with AI?

I am interested in Applied AI and AI Product Engineering opportunities, especially on lean teams where research, engineering, product judgment, commercial understanding and strategic systems thinking need to work together. My technical work is the core; growth and foresight give me additional lenses on users, adoption, incentives and long range consequences.

22f2001536@ds.study.iitm.ac.in

linkedin.com/in/siddharth umathe

GitHub

github.com/output9

Location

Nagpur, Maharashtra, India

Growth work

siddharthumathe.com/work

Download CV

Message sent. I will be in touch soon.

Something went wrong. Please email me directly.

SiddharthUmathe

Systems first. Evidence over adjectives.

AI Powered Software Engineering System

Problem

My role

System modules

Engineering depth

Outcome

What changed in my thinking

Automatic Speech Recognition Systems

Problem

Approach

Engineering depth

Learning

Google Gemma Fine Tuning with LoRA and PEFT

Problem

Approach

Engineering depth

Learning

4× Image Super Resolution

Problem

Approach

Engineering depth

Learning

GAN Style Generative Image Modeling

Problem

Approach

Engineering depth

Learning

Business Data Management · Native Chefs

Problem

Approach

Focus areas

Learning

Applied AI is an engineering problem, not a demo problem.

Built around systems, not keyword lists.

I understand what happens before and after the model.

Product and Positioning

Performance Creative

Paid Acquisition

Growth Automation

Technical depth with a strong applied AI bias.

Signals that matter beyond self description.

Selected academic training.

Problems I want to work on.

Systems thinking beyond the product.

Analytical Lens

Research Themes

Working Standard

Building something difficult with AI?

Siddharth
Umathe