Skip to content

Hello, I'm

Varad Patil

AI/ML Engineer • Researcher

M.Tech student in AI for Sustainability at IIT Kanpur. I build AI systems that go from research to deployment: multi-agent GenAI platforms, retrieval-augmented generation, object detection, and vision-based air-quality sensing for my thesis.

View ResearchResumeFull CV
Full-stack work? See my Dev Portfolio →
Varad Patil, AI/ML Engineer and Researcher

About Me

I am an AI/ML engineer and researcher working across large language models, computer vision and climate data. My thesis at IIT Kanpur estimates PM2.5/PM10 from a city-scale CCTV network, and my projects range from a Gemini-powered multi-agent civic platform to monsoon and cyclone forecasting. I care about rigorous evaluation: leakage-controlled splits, honest baselines and ablations.

Achievements

  • Reliance Foundation Postgraduate Scholarship, one of 100 awarded nationwide (2025)
  • Gold Medalist, Rank 1 in B.Tech Computer Engineering, RIT (2025)
  • Top 20 of 15,000+, Vibe2Ship National Hackathon; Finalist, CityLens AI Hackathon, IIT Kanpur (2026)
  • First-author publication, ICTCS 2024 (Springer)

Positions of Responsibility

  • Teaching Assistant, Mathematics and Computation Using Python, IIT Kanpur (2026)
  • Community Lead, Google Developer Groups on Campus: 45+ members, 5+ workshops on Web & GenAI (2024–25)

Certifications

  • Generative AI with Large Language Models, DeepLearning.AI & AWS (2024)
  • Supervised Machine Learning: Regression and Classification, DeepLearning.AI & Stanford (2025)

Education

2025 – 2027

M.Tech in AI for Sustainability

Indian Institute of Technology, Kanpur

CPI: 9.25/10

Focus: Machine Learning, Deep Learning, Climate & Sustainability AI

2021 – 2025

B.Tech in Computer Engineering

Rajarambapu Institute of Technology, Maharashtra

CPI: 9.70/10 · Gold Medalist (Rank 1)

Focus: Computer Science, Software Engineering

Research & Publications

Vision-Based PM2.5/PM10 Estimation from a City-Scale CCTV Network: architecture diagram

M.Tech Thesis · IIT Kanpur · Ongoing (2026)Ongoing

Vision-Based PM2.5/PM10 Estimation from a City-Scale CCTV Network

Advisor: Prof. Sachchida Nand Tripathi, Dean, Kotak School of Sustainability

Estimating particulate matter concentrations from fixed-camera CCTV footage, pairing clips with co-located reference PM2.5/PM10 monitors and CPCB CAAQMS meteorology on a common observation grid.

Key Contributions

  • •Physics-grounded features (RMS contrast, dark-channel prior, Koschmieder extinction, optical flow) alongside frozen DINOv2 ViT embeddings
  • •Per-camera ResNet-18 regressors benchmarked against a gradient-boosting ladder attributing contribution across optics, embeddings and meteorology
  • •Leakage-controlled protocol with day-, week- and camera-blocked splits and paired blocked-bootstrap significance testing
  • •Resumable multi-stage pipeline with atomic manifests, hash-guarded caches and an automated test suite
PyTorchDINOv2ResNet-18Gradient BoostingComputer Vision
A RAG-Based Legal Chatbot Using Llama2 and Advanced Information Retrieval: architecture diagram

First Author · ICTCS 2024, Springer

A RAG-Based Legal Chatbot Using Llama2 and Advanced Information Retrieval

Designed and implemented an end-to-end retrieval-augmented generation pipeline for legal question answering, with FAISS vector indexing and Llama2-7B served via LangChain.

Key Contributions

  • •Corpus chunking, embedding strategies and prompt templates for legal documents
  • •FAISS-based vector search for efficient semantic retrieval
  • •Evaluation framework measuring retrieval precision, latency and response quality across 500+ test queries
  • •Presented at the International Conference on Trends in Computational and Cognitive Sciences (2024)
Llama2LangChainFAISSPythonFlask

Projects

AI/ML work spanning multi-agent GenAI systems, computer vision, and climate and air-quality forecasting.

CivicPulse: Multi-Agent Civic Issue Platform screenshot

Self Project · Vibe2Ship Top 20

CivicPulse: Multi-Agent Civic Issue Platform

An orchestrated multi-agent pipeline for civic complaints: multimodal triage extracting category, severity and hazards, automated routing to municipal authorities with priority and SLA, and before/after repair verification by image comparison. Inference is reserved for perception and judgement, while deduplication, geospatial hotspot prediction and routing stay deterministic. Deployed on Google Cloud Run within free tiers.

Gemini 2.5 FlashMulti-AgentNext.jsTypeScriptFirestoreCloud Run
LiveCode
CityLens: Urban Infrastructure Monitoring screenshot

Self Project · CityLens AI Hackathon Finalist

CityLens: Urban Infrastructure Monitoring

YOLO11m trained on the MBDD2025 UAV dataset (14.5K images, 57.6K annotations) to 0.918 test mAP@0.5, using iterative stratification and rare-class oversampling for a 10:1 imbalance, trained via DDP on dual T4 GPUs. An illegal-billboard detector reached 0.813 mAP@0.5 through a five-variant ablation with a from-scratch AP metric; SAHI tiled inference showed dataset scale, not architecture, was the binding constraint.

YOLO11PyTorchUltralyticsSAHIDDP
Damage modelBillboard model
Resume Insight: LLM Resume Evaluation Engine screenshot

Self Project

Resume Insight: LLM Resume Evaluation Engine

A Gemini-backed analysis service over PDF-extracted resume text with separate endpoints for job-match scoring, ATS analytics and section-level rewriting, with per-request token and error logging. Semantic resume-to-JD matching, keyword alignment and line-level suggestions, deployed publicly with tiered subscriptions and per-feature quotas on Clerk and Supabase.

Gemini APILLMsNext.jsClerkSupabase
LiveCode
Active-Break Indian Summer Monsoon Forecasting screenshot

Course Project · KSS605

Active-Break Indian Summer Monsoon Forecasting

A TCN–Transformer meta-ensemble classifying Active, Break and Normal monsoon phases up to 7 days ahead from ERA5 reanalysis. Lifted mean Macro F1 to 0.560 against an XGBoost baseline at 0.521, with Break-class F1 at the 7-day horizon rising from 0.198 to 0.493. Labels come from IMD rainfall anomalies with training-only thresholds and chronological splits to prevent leakage.

PyTorchTCNTransformersXGBoostERA5
Code
Cyclone Track Prediction System screenshot

Course Project · EE708

Cyclone Track Prediction System

Multi-horizon cyclone track forecasting (6h to 72h) from IBTrACS with 40+ engineered motion, intensity and spatial features. Achieved 14.47 km RMSE at 6h and 78.77 km at 24h with geodesic error metrics and skill scores against a persistence baseline. Persistence-extrapolated positions act as a physics anchor, and storm-level splits prevent track leakage.

XGBoostGradient BoostingScikit-learnIBTrACS
Code
Air Quality Early Warning System screenshot

Self Project

Air Quality Early Warning System

Next-day AQI forecasting for Indian cities with paired regression and six-class CPCB bucket classification, trained on 108K station-day records across 110 stations, reaching R² 0.92 and 20.99 AQI MAE on held-out data. Recursive seven-day forecasts are served through a Flask REST API with a dashboard of AQI trends and health advisories.

Gradient BoostingScikit-learnTime SeriesFlaskPython
Code
More full-stack & web projects (Snikrz, Spotify 2.0, DJB WhatsApp CRM) on my Dev Portfolio

Experience

Altair Engineering logo

Altair Engineering

Jan 2025 – Jun 2025

Software Development Intern · Bangalore

  • •Built full-stack MERN client portals for tire manufacturers to configure, submit and track HPC simulation workflows
  • •Built a real-time HPC job scheduling and visualisation dashboard surfacing live job status, queue position and results
  • •Improved frontend performance by 30% through code-splitting, lazy loading and API caching
  • •Automated simulation and reporting workflows with Python and REST APIs, cutting manual operations by 40%
Bubble Byte Ventures logo

Bubble Byte Ventures

Sep 2022 – Aug 2023

Web Development Intern · Sangli

  • •Delivered 10+ responsive, SEO-optimised client websites, reducing page load times by 30%
  • •Automated deployment pipelines with Python
  • •Received the Best Performance Award for consistent delivery across client projects

Skills & Technologies

GenAI & LLMs

RAG SystemsLangChainFAISSLlama2GeminiOpenAI APIHuggingFacePrompt EngineeringMulti-Agent SystemsMultimodal Inference

Deep Learning & Vision

PyTorchTensorFlowTransformersTCNCNNsObject Detection (YOLO, Ultralytics)

Machine Learning

Scikit-learnXGBoostGradient BoostingTime-Series ForecastingFeature EngineeringImbalance HandlingAblation & Baseline Analysis

Scientific Data

NumPyPandasMatplotlibERA5 ReanalysisIBTrACSCPCB / CAAQMS Air Quality

Languages & Backend

PythonTypeScriptJavaScriptFastAPIFlaskNode.jsREST APIs

Cloud & Data

GCP (Cloud Run, Firestore)DockerMongoDBPostgreSQLSupabaseGit

Get in Touch

I'm always open to discussing research collaborations, AI/ML opportunities, or interesting projects. Feel free to reach out!

Download ResumeFull CV

Email

varadapatil123@gmail.comvaradap25@iitk.ac.in

LinkedIn

Connect with me

GitHub

View my code

Dev Portfolio

Full-stack & web work