Pratham Garg — builds ML things, runs hills, organizes chaos

01 / about

Who's this guy?

I'm Pratham. I study computer science at Thapar, and most days you'll find me somewhere between a half-finished Jupyter notebook and a gym session.

The honest version: I only made my GitHub account in May 2025. Since then it's grown into 26 repos, a package on PyPI, and a seat at Amazon ML Summer School — 80,000 people applied, about 3,000 got in, and I still think about that email. I like building things that come from real life: a stress detector that reads your body's signals, a waste classifier for smarter cities, and a café finder I built because a trek to Jibhi left me hungry and mildly annoyed.

I'm not the "started coding at 8" type. I'm the type that shows up every day — LeetCode streaks, 30 Nites of Code, more certificates than I'll ever list. Outside the editor I run, lift, trek, and have helped organize fests for 5,000+ people. One season I was even an actor and model. Long story — ask me.

Right now: third year, looking for a Summer 2026 internship where I can ship real things alongside people better than me.

02 / work

Things I've actually built

Every one of these started as a real problem — mine or somebody's. Straight from the READMEs, no marketing translation layer.

01

Yogic

Can a model tell you're stressed before you can?

An AI system that detects and quantifies stress from physiological sensor data — EDA, BVP, heart rate, SpO₂. It uses a Mixture-of-Experts deep network where different experts learn different sensor patterns, and a gating network fuses them into a continuous stress score plus a neutral / non-neutral call. Evaluated with F1 and AUROC, tested with real-time inference.

the honest part ↓

What was hard: four sensors, four different sampling realities. Getting the streams aligned and the experts to actually specialize took longer than the model itself.

Next time: try attention-based encoders per sensor and a proper ablation of the gating network.

PythonTensorFlowMixture of Expertstime-seriesbiosignals

repo ↗

02

The Pickleball Arena

A real booking site, for a real business, in my hometown.

A court-booking platform for a pickleball facility in Bathinda, Punjab — two LED-lit championship courts, online reservations with date/time slots, membership tiers, tournament sign-ups, and WhatsApp-first contact because that's how people in Bathinda actually book things. Not a portfolio demo: actual players reserve actual courts on it.

the honest part ↓

What was hard: the users aren't tech people — they're players booking a court from their phone between games. Every extra tap costs a booking.

Next: the coaching module. The site says "launching soon," so I should probably launch it.

live for a real businessbookingsVercelBathinda represent

live ↗ repo ↗

03

Happy Hoping

Born on a trek. true story

I trekked to Jibhi, Himachal. Loved it. Couldn't find a single decent café recommendation online. So instead of just complaining, I came home and built the fix.

“We've all been there — staring at a list of restaurants, unable to decide where to eat. Decision fatigue is real... I wanted to build a tool that feels less like a directory and more like a smart friend who knows your taste.” — my own README, and I stand by it

A personalized food-discovery engine that learns your preferences instead of pushing sponsored lists. Started with Jibhi; still growing. Built in public on LinkedIn.

the honest part ↓

What was hard: small-town India barely exists in food-data APIs. That gap is literally why the project exists.

Next time: a real preference model instead of heuristics, and more towns.

JavaScriptrecommendationsshipped & live

live ↗ repo ↗

04

CogniAI

Teaching cameras to sort a city's garbage.

Computer-vision waste classification for smart-city recycling — built with my teammate Sanyam Wadhwa. Real-time prediction, preprocessing pipelines, inference tuned for efficiency, and modular components meant for actual deployment rather than a notebook demo.

the honest part ↓

What was hard: "accurate" and "fast enough for real-time" pull in opposite directions. Most of the work was in the middle.

Next time: edge deployment and a feedback loop from misclassifications.

Pythoncomputer visionCNNssustainability

repo ↗

05

TOPSIS, on PyPI

My first published package. Roll number included.

pip install Topsis-Pratham-102303052 — a working implementation of TOPSIS, the multi-criteria decision-analysis method, published to the real Python Package Index with versioning, docs and all. Also applied it to a text-classification model-selection study. Naming a package after your roll number is peak engineering-student energy and I refuse to apologize for it.

the honest part ↓

What was hard: the method took an afternoon; packaging, metadata and versioning took the weekend.

Next time: bundle more MCDM methods under one clean API.

PythonPyPIMCDMpublished

PyPI ↗ repo ↗

more experiments —

transformer — flagging Indian grid cells that need distribution transformers; CLI + folium maps
numerai — tournament ML experiments
gdp-dashboard — Streamlit data app
campuspo — Next.js, in progress
neetcode-submissions — the daily grind, in C++

03 / proof

Don't take my word for it

Pulled live from the GitHub API every time you load this page.

0public repos

0days since first repo

0languages

0certifications

jupyter · html · python · typescript · c++ · other

recently pushed —

loading from GitHub…

Account created May 2025. Yes, everything above happened in one year. Imagine year two.

GitHub ↗ Kaggle ↗ LeetCode 50-day badge ↗

04 / timeline

How I got here

2011–21
St Xavier's, Bathinda. Cabinet member. The roots.
Aug 2023
Started B.Tech CSE at Thapar. Joined SSA VIRSA's executive committee within weeks — stayed a full year.
2023–24
VIRSA cultural fest. 15-person core team, 5,000+ attendees.
Nov 2024
Saturnalia fest — core member and actor/model. ★ Yes, really.
Jan 2025
Executive committee, URJA TIET.
Mar 2025
Year Representative, MUDRA Society — the voice for 200+ students.
May 2025
Created my GitHub account. first-repo, in C. Humble beginnings.
Jun 2025
Deloitte Australia cyber simulation (Forage).
Aug 2025
Selected for Amazon ML Summer School '25. 80,000+ applied. ~3,000 selected.
Oct 2025
Dynamic Programming Camp, AlgoUniversity.
Jan 2026
Published TOPSIS to PyPI. Also: Google AI Essentials.
Jan–Feb 2026
Built CogniAI and Yogic.
Jun 2026
LeetCode 50-Day Badge · 30 Nites of Code in progress. The grind continues.
Now
You, hopefully →

credentials that matter —

Amazon ML Summer School '25 NVIDIA · Fundamentals of Deep Learning IBM · Generative AI Google AI Essentials AlgoUniversity · DP Camp

29 total — 9× AI/GenAI · 7× SQL & data · 5× ML/DL · +24 more on LinkedIn ↗

05 / skills

What I work with

daily drivers

PythonPandas / NumPyScikit-learnGit & GitHubJupyterSQLC++ (DSA)

building with

TensorFlowPyTorchFastAPIDockerMySQL / PostgreSQLBigQueryREST APIsAWS (basics)

actively learning

LLM APIs & RAGvector DBs (FAISS / Chroma)Next.js / TypeScriptMLOps

Tiers, not percentages — nobody knows what 80% of Python means.

06 / beyond code

The rest of me

🎧 On loop

popR&BPunjabiHindi

🏃 Hybrid athlete

run · lift · trek · repeat

Progressive overload works on muscles and on ML skills. One trek even shipped a product →

🎪 The fest life

Organizing for 5,000 people teaches you more about systems than any textbook. VIRSA exec committee, one full year.

🎬 Stage & camera

Saturnalia '24 — actor and model, one fest season. Yes, really. Ask me about it.

📍 Bathinda → Patiala

Punjab raised. The marigold and tangerine on this page aren't an accident.

💬 Ask me about

Jibhi treks · Mixture-of-Experts models · fest-night chaos · the acting stint · why my PyPI package has my roll number in it.

07 / now

Currently

→ studying : 3rd year CSE, Thapar

→ building : this site · campuspo

→ grinding : 30 Nites of Code · post-50-day-badge LeetCode streak

→ training : run · lift · planning the next trek

→ looking for : Summer 2026 AI/ML internship

last updated · June 2026