Mayank Goley

01. ABOUT ME

Hi, I'm Mayank. Thanks for stopping by.

I'm a software engineer who enjoys building AI products that people actually use. Over the past four years I've worked at Capgemini, Infosys, and most recently Acento Lujo, where I built an AI assistant that took a real load off the customer support team. Along the way I've learned that good AI work is less about chasing the newest model and more about building something reliable, affordable, and genuinely helpful.

Right now I'm finishing my Master's in Computer Science at Cal State Fullerton and graduate in December 2026. For my capstone I built HireNvy, a tool that helps job seekers see how well their resume fits a role, and I spend much of my free time on projects that explore how to make AI more trustworthy.

I'm looking for AI Engineer, ML Engineer, and Software Engineer roles with teams that care about building AI thoughtfully and shipping it well.

Outside of work, you'll usually find me watching baseball, cooking for friends, or at a tech meetup.

02. WORK EXPERIENCE

Software Engineer (AI/ML)

Jul 2023 – Jul 2024

Acento Lujo | Ghaziabad, India

Resolved about 30% of tier-1 customer questions with a RAG support assistant (GPT-3.5 Turbo, pgvector, metadata-filtered retrieval) over the 130-SKU catalog, shipping and return policies, and past support tickets, handing the rest to staff.
Held LLM costs under $150 a month via response caching and prompt trimming, and guarded answer quality with a 100-question test set from real customer chats, run before every prompt or model change.
Eliminated 10+ hours/week of sales reconciliation via a Python/PostgreSQL ETL pipeline for Shopify and Amazon India orders.
Cut Diwali 2023 stockouts about 18% vs. 2022 with XGBoost demand forecasts trained on 18 months of sales and lead times.

Software Developer

Nov 2022 – Jun 2023

Infosys (Client: Banking client, biometric fraud prevention) | India

Reached sub-300ms latency on a real-time liveness detection system for a banking client's biometric fraud prevention through model optimization and INT8 quantization.
Made model deployments reproducible with MLflow tracking, Docker packaging, and Terraform infrastructure as code.

Software Engineer

Feb 2020 – Jul 2022

Capgemini (Client: Healthcare client, digital patient portal) | India

Reduced integration errors by 12% on a healthcare client's patient-portal Flask APIs by tightening request validation and structured JSON handling, backed by pytest tests.
Shrank bundle sizes by 15% and first-load times by 5% with a custom Python/Flask CSS minification utility.
Delivered RESTful API endpoints for the client's digital patient portal end to end using Flask, SQLAlchemy, and PostgreSQL.

03. PROJECTS

HireNvy

My M.S. capstone. Scores resumes three ways against 1,300+ job descriptions scraped from Greenhouse, Lever, Ashby, SmartRecruiters, and Workday: formatting checks, O*NET/ESCO role fit, and SBERT embedding match in Qdrant. A DistilBERT model I fine-tuned tags resume sections at 97.3% macro F1 on 1,000 held-out resumes, and match scores reach 78% pairwise ranking agreement with recruiter reviews. Built as separate Django, FastAPI, and Next.js services linked by Dramatiq queues and JWT auth, and piloted by 45 peers in the department.

FastAPI, Django, Next.js, PyTorch, Hugging Face, Qdrant

LLM Evaluation and Observability Framework

A Python SDK that logs every LLM call and grades responses asynchronously at under 10ms overhead per call. The LLM judge is calibrated on 150 hand-labeled cases to 0.73 Cohen's kappa with human ratings, and a GitHub Actions quality gate fails pull requests when answer quality on the Financial Intelligence System's golden finance dataset drops below threshold.

Python, GPT-4o-mini, GitHub Actions, LLM-as-a-Judge

Financial Intelligence System

Streams cited financial reports from LangGraph Bull, Bear, and Synthesis agents over up to 10 SEC filings per ticker via FastAPI, with hybrid Qdrant plus BM25 retrieval. A 3-stage verifier (fiscal year alignment, citation mapping, data grounding) rejects about 12% of generated claims as unsupported and retries up to 2 times.

FastAPI, LangGraph, Qdrant, BM25, Claude

TeachWise

A full stack tutoring marketplace connecting students with tutors. Features real time video sessions, geo based search, and split payment processing. Built with Flask, Next.js 14, PostgreSQL with PostGIS, Stripe Connect, and Jitsi.

Flask, Next.js 14, PostgreSQL, PostGIS, Stripe Connect, Jitsi

Ask

An offline first desktop RAG app I'm building with Tauri and Rust. Retrieval augmented generation that runs locally, with no network connection needed.

Tauri, Rust, RAG

Energy Forecasting

Deep learning model forecasting 24 hours of energy demand on 16 years of hourly PJM grid data. Built with PyTorch, comparing LSTM and Transformer architectures.

PyTorch, LSTM, Transformer

Traffic Sign Classification

Multi class classifier for German traffic signs across 43 categories. Built with PyTorch, comparing a scratch CNN against a frozen ImageNet pretrained ResNet18.

PyTorch, CNN, ResNet18

Speech Emotion Recognition

Eight class speech emotion classifier on the RAVDESS dataset using log mel spectrograms. Built with PyTorch and librosa.

PyTorch, librosa

Face Mask Detection

Image classifier that detects whether a face is wearing a mask. A CNN trained from scratch reached 0.957 accuracy, and a frozen pretrained MobileNetV2 reached 0.982.

PyTorch, CNN, MobileNetV2

IMDB Sentiment Analysis

Classifies movie reviews as positive or negative, comparing TF-IDF with Logistic Regression (0.905 accuracy), an LSTM (0.862), and DistilBERT (0.913).

scikit-learn, LSTM, DistilBERT