Irish Mehta | Machine Learning Engineer, Applied Scientist & Researcher

Summary

A client-side Sudoku learning app with real-time game state, hints, image import, OCR review, and candidate-note tools.

Summary

An automated job-discovery system that finds, ranks, and publishes ML and data opportunities, with tailored-resume generation for top matches.

Problem Statement

Searching company-by-company was repetitive and made it difficult to consistently identify roles compatible with location, seniority, and work-authorization constraints.

Methodology

• Scraped and normalized jobs from seven ATS families with throttling, retries, and title/location filtering.
• Used a two-pass LLM pipeline for fit analysis, structured scoring, and eligibility gates.
• Published curated roles to a searchable board and generated tailored LaTeX resumes for highly ranked opportunities.

Results

• Aggregated 412K+ postings across 9K+ companies and identified 3.2K relevant ML/data roles.
• Reduced tailored-resume generation to under two minutes per role.

Summary

An autonomous photography assistant that iteratively improves portrait framing by reasoning about a scene and moving a physical camera rig.

Problem Statement

Getting a well-composed shot requires repeated manual camera adjustments and visual judgment.

Methodology

• Ran a vision loop: capture an overview, retrieve similar examples, propose a pose, capture a candidate, score it, and iterate.
• Used Gemini for scene understanding and scoring, an Android phone for capture, and an ESP8266-controlled rig for pan, height, orientation, and zoom.

Results

Autonomously iterates toward stronger framing for up to 15 steps, stopping when the score reaches a threshold or improvement stalls.

Summary

A multi-sensor foundation model that unifies Mars orbital imagery from HiRISE, CTX, and THEMIS for downstream remote-sensing tasks.

Problem Statement

Martian sensors capture imagery at dramatically different resolutions, making direct combination of their representations inefficient.

Methodology

• Pre-trained masked autoencoders on roughly 12 million images from the three sensors.
• Merged models using Equal Validation Loss and task arithmetic.
• Built and parallelized downstream evaluation across nine Mars-Bench tasks and strong Earth-observation baselines.

Results

MOMO outperformed ImageNet and Earth-observation baselines, averaging about a 1% mIoU gain on segmentation tasks and up to 4% on the Boulder dataset.

Summary

A standardized benchmark of 20 Mars-vision datasets spanning classification, semantic segmentation, and object detection.

Problem Statement

Martian imagery was difficult to compare across studies because datasets, task definitions, and evaluation conventions were fragmented.

Methodology

• Built a scalable PyTorch benchmarking pipeline and evaluated 13 CNN, Transformer, and Earth-observation model families.
• Tested training from scratch, full fine-tuning, and frozen feature extraction across standard, partitioned, and few-shot splits.

Results

Completed 19K+ model runs and 780 hyperparameter sweeps on ASU HPC infrastructure, establishing reproducible baselines for the benchmark.

Summary

An adaptive learning system that adjusts question difficulty in real time using Multi-Armed Bandit algorithms.

Problem Statement

Fixed-difficulty practice does not respond to an individual learner's changing ability or distinguish long-term learning from a one-off assessment.

Methodology

• Implemented a persistent GradientBandit learning mode and a session-only Adaptive Epsilon-Greedy testing mode.
• Modelled learner ability across four difficulty levels and used cached, topic-filtered questions to avoid repetition.

Results

Provides persistent ability tracking and recommendations for learning, while keeping test sessions independent and resettable.

Summary

A food-discovery app that lets people search restaurant menus in natural language and filter by dietary needs, allergens, macros, and price.

Problem Statement

Menu information is unstructured and makes it difficult to find a meal that satisfies several constraints at once.

Methodology

• Extracted structured dish data from PDFs and images using a multi-document menu-analysis pipeline.
• Converted natural-language requests to Snowflake SQL with Gemini, using Cortex Analyst as a fallback.
• Delivered results through a React frontend and FastAPI service with caching and rate limiting.

Results

Supports complex multi-constraint searches across a Tempe pilot dataset, with cached queries returning in under two seconds.

Summary

A production corporate website and digital product catalogue for an industrial silicone manufacturer.

Problem Statement

The business needed a discoverable, secure, mobile-friendly way to present its product range and capture customer enquiries.

Methodology

• Built the catalogue with Next.js, TypeScript, SQLite-backed product data, and static product pages.
• Connected quote and contact forms to Zoho CRM, with validation, rate limiting, and security headers.
• Implemented structured data, metadata, sitemaps, and optimized AVIF/WebP imagery and video.

Results

• Published 101 products across six categories and generated 118 static pages.
• Reduced hero-video size by 92% and homepage industry-image size by 96%.

Summary

A strategic counter-narrative system that identifies social-media users susceptible to polarization and generates context-aware counter-messages.

Problem Statement

Echo chambers reinforce polarized viewpoints, but a useful intervention needs to account for a user's stance and openness to alternative narratives.

Methodology

• Used a modified HITS algorithm on user-topic graphs to produce vulnerability scores and classify users.
• Detected stance across nine geopolitical dimensions, selected an opposing narrative, and built persona-aware LLM prompts.
• Exposed the workflow through FastAPI and a Flask interface.

Results

Produces a user classification, stance vector, and targeted counter-message from a tweet, user ID, and political-camp context.

Summary

A session-based RAG system for asking grounded questions of earnings-call transcripts and financial documents.

Problem Statement

Financial documents are lengthy and difficult to query quickly, while sensitive uploads require isolation between users.

Methodology

• Parsed uploads with LlamaParse, created semantic chunks, embedded them with HuggingFace models, and retrieved from session-specific FAISS stores.
• Used FastAPI, Streamlit, LangChain, and Groq to generate source-grounded responses with page and speaker references.

Results

Each session has an isolated vector store with automatic cleanup after 30 minutes of inactivity, supporting private, attributable document Q&A.

Summary

A content-based movie recommender that combines metadata, trailer visuals, and audience sentiment to find films with a similar feel.

Problem Statement

Genre and cast alone do not capture the visual tone and audience reception that shape whether two films feel alike.

Methodology

• Used TF-IDF and cosine similarity on keywords, cast, genres, director, and overview to find candidates.
• Extracted trailer key frames, then used VGG19 features and K-means clustering for visual similarity.
• Added a review-sentiment score to create the final composite ranking.

Results

The best sentiment classifier, Linear SVC, achieved 90% accuracy and an 89.85% F1 score on IMDb reviews.

Summary

A distributed IoT system that detects faces on an edge device and performs cloud-based recognition.

Problem Statement

Real-time recognition needs fast local detection without sending every frame through a cloud inference workflow.

Methodology

• Ran MTCNN face detection on an AWS Greengrass edge device.
• Sent recognition requests through AWS IoT Core and SQS to a Lambda service using FaceNet embeddings and cosine similarity.
• Added a 60-second TTL filter to prevent duplicate recognition requests.

Results

Created a deployable edge-to-cloud recognition pipeline with MQTT, queued request/response handling, and Greengrass artifact deployment.

Summary

A private, local analytics platform that imports Gmail, classifies job-search messages, and turns them into an editable application pipeline.

Problem Statement

Manual inbox searching and fragmented tracking make it difficult to see application status, next actions, and recruiter activity.

Methodology

• Incrementally synced Gmail into SQLite and classified messages with cached, versioned Claude taxonomies.
• Resolved company identities, derived company-level pipeline analytics, and supported non-destructive user overrides.
• Presented KPIs, funnels, timelines, aging, and action queues in a local Plotly Dash application.

Results

Processed 13,403 emails into 4,658 job-search classifications, identifying 816 companies applied to and 1,547 distinct roles.

Summary

An evaluation of practical dropout-based methods for quantifying vision-model uncertainty during inference.

Problem Statement

Monte Carlo dropout can require up to 1,000 forward passes per image, making uncertainty-aware inference computationally expensive.

Methodology

• Compared six dropout variants on MobileViT-S trained on CIFAR-10.
• Developed an early-stopping criterion based on convergence of the running mean and variance of prediction probabilities.

Results

Variance-based stopping reduced inference cost by over 95%; variational dropout achieved 89.75% accuracy with about 30 passes instead of 1,000.

Summary

An event-driven personalization system that scores players at the end of a game and selects an appropriate bonus using predicted second-purchase propensity.

Problem Statement

Player incentives need to be timely and individualized while remaining configurable by business teams and consistent with the latest gameplay and wallet data.

Methodology

• Consumed game-end events from RabbitMQ and fetched aggregated player features from a Snowflake pipeline refreshed approximately every five minutes.
• Served a versioned XGBoost model on AWS EC2 using 15 ordered gameplay, wallet, purchase, and withdrawal features.
• Converted propensity scores into low, medium, or high probability buckets, applied treatment rules, and resolved the resulting segment through admin-managed bonus mappings.

Results

Connected event-driven inference, configurable segmentation, and a frontend integration layer into a production flow for near-real-time personalized offers.