AMIT

Hey, I'm Amit,

An AI/ML
Engineer
& On-Device
ML Builder

Shipping transformer models to phones, laptops, and everything in between — where the cloud isn't invited.

Contact Me
Amit Yadav

Field Studies

001 — 001

Small, finished software aimed at problems real companies are staffing teams to solve. One industry at a time, taken far enough to measure and then stopped.

CPG · Trade Finance2026

Shortpay

Distributors pay a brand's invoice short and attach a reason code. The brand has 90 to 180 days to dispute before the money is gone for good, and the documents proving it arrive as scans with no text layer. Shortpay reads them, reconciles every deduction against the deal sheet and delivery paperwork, and prepares the dispute.

Most tools rank disputes by dollar size. That is the wrong sort — a $400 claim closing on Friday is worth more than a $4,000 claim with 90 days left, because only one of them is about to become zero.

What it measured
90.6% recall at 93.9% precision against planted defects, 98% field accuracy on scanned documents, and one defect class left at zero because the data to solve it does not exist.
Built with
Claude (Vision)
FastAPI
React
TypeScript
Open itSource
On-Device Inference

CoreML, quantization, Swift tokenization.

CoreML
Swift
Hugging Face
LLM Fine-Tuning

Unsloth, domain datasets, MLOps loop.

PyTorch
Unsloth
Hugging Face
Full-Stack ML

Data pipelines → API → AWS deploy.

AWS
Docker
FastAPI
Open Source

sktime, Kiwix/openZIM, upstream contributions.

GitHub
sktime
Kiwix

Building ML that runs where people actually are.

I work on the seam between Hugging Face and constrained hardware — closing the gap between a 7B model on a research cluster and a quantized transformer running inside an iPhone. That throughline connects everything I build: Cortex, my on-device LLM research at BU, and the Gutenberg semantic-search work I'm proposing for GSoC 2026. When the cloud isn't available — or isn't welcome — the model still has to be there.

Selected Work

001 — 003
Cortex
SwiftCoreMLDistilBERTSwiftData
iOS · On-Device

Cortex

A privacy-first iOS bookmark manager where a DistilBERT model auto-tags links fully on-device — no accounts, no round-trips, no telemetry. The trickiest part was making the model optional: if the weights aren't bundled, the app falls back to Apple's NaturalLanguage framework so tagging still works on a slim build. SwiftData handled persistence almost invisibly, which let me spend the time on the part I actually cared about — tuning the tag vocabulary until the suggestions felt like mine.

Built with
iOS
Swift
CoreML
Hugging Face
Xcode
Source
Smart Energy Optimizer
SwiftUIIBM watsonxGranite-3WebSockets
IBM TechXchange Hackathon · 2025

Smart Energy Optimizer

Built in a weekend for the IBM TechXchange hackathon: a SwiftUI dashboard that streams live telemetry over WebSockets from a Node/Express backend wrapping IBM watsonx's Granite-3-8b for 24-hour load forecasting and peak-shaving tips. The backend fakes a house full of HVAC and lighting through a simulator, which turned out to be the right call — we could demo the whole loop without anyone rewiring an apartment. I pushed the socket layer to a 30-second heartbeat so the SwiftUI side could stay purely reactive via @Observable instead of babysitting polling timers.

Built with
Swift
iOS
Node.js
JavaScript
Docker
Source
Modern Data Warehouse
PostgreSQLPL/pgSQLMedallion ETLTableau
Data Engineering

Modern Data Warehouse

An end-to-end ETL pipeline in pure PostgreSQL and PL/pgSQL — no Airflow, no dbt, no managed warehouse. Raw sales data moves through a Bronze → Silver → Gold medallion architecture and lands in a star schema ready for Tableau. It's the least flashy project in my portfolio and the one that taught me the most about why the boring parts of data engineering — keys, grain, idempotent loads — are the parts that actually matter.

Built with
PostgreSQL
Source

Entrepreneurship

All Stories
Entrepreneurship · Harvard GovLab Pitch Day
Beyond Policy and Code: What I Learned at Harvard's GovLab

Attending the GovLab pitch day at Harvard Kennedy School showed me that the most exciting ideas aren't just about technology or policy—they're about people bridging the gap between them.

MAY 12, 2026

Writing

All Posts
Web Development
Making GitHub Contributions Visible — Bringing Your Open Source Activity to Life on Your Portfolio

How I built a live GitHub contribution graph with automatic PR status tracking, pulsating indicators for significant contributions, and clickable links to GitHub. A technical walkthrough of fetching, visualizing, and animating GitHub API data for recruitment and community impact.

MAY 11, 2026
Open Source
GSoC 2026 Rejection: Why Semantic Search for Kiwix Didn't Happen

I proposed semantic search for offline book discovery to Kiwix in GSoC 2026. It was rejected. This is why, what I learned, and why organizations are right to sometimes say no to new ideas.

MAY 05, 2026
AI & Security
From 213 Vulnerabilities to Project Glasswing — What Three Years Did to AI-Powered Security

In 2023, GPT-3 found 213 security flaws by scanning files one at a time. In 2026, Claude Mythos Preview finds thousands of zero-days in every major operating system, autonomously. The distance between those two data points should keep you up at night.

APR 13, 2026