Skip to content
MG / build log
Résumé, one page PDF

Mehul Gupta — build log

Move the cursor across the portrait to develop it

MEHULGUPTA

Mehul Gupta, in a black suit.

Open to

  • Software Development Engineer
  • Backend Engineer
  • Full-Stack Developer
  • Machine Learning Engineer
  • Product Engineer

I build systems people trust with things that matter.

A payment, a ride home, a season's harvest. Final-year CSE at Bennett University — building products end-to-end at Sahayogi One, and researching machine learning for precision agriculture.

See what I've shippedRead how I think
Now
Developer Trainee, Sahayogi One
Studying
B.Tech CSE, Bennett University
Research
Hyperspectral ML for crop yield
Base
Karnal / Greater Noida, IN
§02Revision history

What broke, and what replaced it

The case files show the systems and the roles show the work. This is the part neither can show: 7 revisions, 5 of them breaking. A major bump means something I was building on turned out to be wrong.

  1. v0.12023

    Initial commit

    B.Tech CSE, Bennett University.

    • Added C++, data structures and algorithms — the cost model before the frameworks.

  2. v1.0Feb 2025Major

    I had scoped the wrong problem

    Ocasio.

    • Breaking I had scoped it as a scheduling problem. It was a trust problem — verification and payment integrity, not the booking flow.

    • Added First end-to-end build: listings, recommendations, payments.

  3. v2.0Apr 2025Major

    The framework is not the floor

    CUDA C with NVIDIA; operating systems with Google.

    • Breaking I could describe what my code did and not what the machine did. Those are different claims, and I had been making the wrong one.

    • Added Parallelism at the metal, and the layer scheduling it.

  4. v3.0Jun — Jul 2025Major

    Working and keepable are different properties

    RBH Solutions — first professional codebase.

    • Breaking My definition of done was 'it runs'. In a codebase a team inherits, that is roughly the halfway mark.

    • Removed Writing code whose only reader was me.

  5. v3.1Mar 2026

    The same lesson, applied on purpose

    UniRyde.

    • Added Verification designed in as the entry condition, rather than discovered halfway like it was on Ocasio. Nothing broke here — that is the whole point of the entry.

  6. v4.0Apr 2026Major

    A single split was never a result

    Hyperspectral yield research.

    • Breaking I had been reading a single split as a result. Across environments those numbers do not survive, which makes the headline figure the least interesting one.

    • Added RFECV, a model bench rather than a favourite, and cross-environment evaluation.

  7. v5.0Jun 2026 — presentMajor

    The spec is an output, not an input

    Sahayogi One.

    • Breaking I had treated the spec as the input to engineering. It is an output of talking to whoever has the problem — and it moves.

    • Added Shipping where a platform's rules and a data-protection law are hard constraints, not preferences.

§03System map

What I work with, in the order a request travels

A list of logos tells you nothing about whether someone can hold a system in their head. This is the same set, arranged the way it actually connects — and labelled honestly by how far I've taken each piece.

InterfaceIntegrationLogicDataPlatform

Interface

What the user actually touches

  • React.jsShipped

    UniRyde's front end — routing, state, map views.

  • JavaScriptShipped

    The glue across both platform projects.

  • HTML / CSSShipped

    Hand-written layouts before reaching for a kit.

Integration

Where my system meets someone else's

  • Google Maps APIShipped

    Routing and pickup matching in UniRyde.

  • REST APIsShipped

    The contract between client and service.

  • Auth flowsShipped

    Student verification and secure sign-in.

Logic

Where the decisions get made

  • PythonShipped

    Research pipeline and automation work.

  • C++Learning

    The language I learned data structures in.

  • JavaLearning

    Coursework and OOP fundamentals.

  • DSAShipped

    Not a checkbox — how I reason about cost.

Data

What has to still be true tomorrow

  • MongoDBShipped

    UniRyde's document store — users, rides, matches.

  • SQLLearning

    Relational modelling and querying.

Platform

What it runs on

  • AWSLearning

    Cloud fundamentals — deployment is where projects go to die.

  • Operating systemsLearning

    Certified with Google, Feb 2025.

  • CUDA / GPULearning

    NVIDIA CUDA C — parallelism at the metal.

Intelligence — branch

The branch that decides under uncertainty

  • XGBoostResearch

    Gradient boosting for yield regression.

  • SVRResearch

    Support vector regression baseline.

  • ElasticNetResearch

    Regularised linear baseline on spectral features.

  • 1D-CNNResearch

    Deep model over the spectral axis.

  • Stacked ensemblesResearch

    Where the accuracy actually came from.

  • RFECVResearch

    Feature selection across a very wide spectral space.

  • GANsLearning

    DeepLearning.AI, Apr 2026.

  • Agentic automationLearning

    Clawdbot — automating my own workflow first.

23 entries

§04Experience

Products in production

Work shipped for a company, where the constraints are somebody else's business, a regulator, and a platform that can shut you off.

Sahayogi One

Developer Trainee
Jun 2026 — present

Working across business verticals — product development, customer interaction and solution delivery — on two products built for Indian SMEs.

BoSS — Business Operations & Systems Suite

Launched

A business operating system for Indian SMEs, spanning 18 modules across the transactions a small business runs on.

Built the People & Payroll and Travel & Expense modules.

NestJSNext.js
Product site, BoSS — Business Operations & Systems Suite

Chat with Sahayogi

A WhatsApp Business Platform product for Indian SMEs — bulk messaging inside Meta's rules and India's consent law.

Worked across the product, with a focus on WhatsApp Business API onboarding.

NestJSNext.js
Product site, Chat with Sahayogi

RBH Solutions Private Limited

Full Stack Developer Intern
Jun — Jul 2025

First professional codebase. Enterprise software under industry-standard development practices, and the difference between code that works and code a team can keep.

§05Case files

Three problems, and what I decided

Screenshots show what a thing looks like. These show what it cost to decide — the constraint, the call I made, and what I gave up to make it.

CF-01

Ocasio

Marketplace platform · Feb 2025Built

Verified vendors, recommendations and payments for an industry that runs on forwarded contacts.

Web platformRecommendationsPayments

Problem

Event planning runs on forwarded contacts and word of mouth. Organizers, photographers and planners are hard to verify, quality is unknowable in advance, and money changes hands on nothing but trust.

Approach

  1. 01Verified listings, so the directory is an assertion rather than a phone book.
  2. 02AI-powered recommendations to match a planner to an event they haven't run before.
  3. 03Secure payments, so the transaction isn't the weakest link in the chain.
How it fits together

Organizer

lists a service

Verification

identity + credentials

Catalogue

verified listings only

Recommendations

planner ↔ event fit

Payments

secure settlement

Role

Concept, product definition and implementation.

The trade-off

[DRAFT] The scarce effort went to verification and payment integrity rather than the booking interface — a beautiful booking flow on top of unverifiable vendors solves the wrong half of the problem.

Private repo

CF-02

UniRyde

Ride-sharing platform · Mar 2026Built

Student-verified ride sharing, where trust is the entry condition rather than a profile badge.

React.jsHTMLCSSMongoDBGoogle Maps API

Problem

Students travelling identical routes each pay for a whole cab. General ride-share apps can't prove the person you're matched with belongs to your university — which is exactly the assurance that makes sharing acceptable.

Approach

  1. 01Student verification as the entry condition, not an optional profile badge.
  2. 02Secure authentication protecting who is inside the matching pool.
  3. 03Trusted matching and route logic over the Google Maps API.
How it fits together

Student

university sign-up

Verification

the entry gate

Auth

session + identity

Matching

route overlap + trust

Maps API

route, pickup, ETA

Shared ride

cost split

Role

Full-stack build — front end, data model and integrations.

The trade-off

[DRAFT] Constraining the pool to verified students shrinks liquidity — fewer possible matches — in exchange for a pool where every match is defensible. For a safety product that trade is worth making.

Private repo

CF-03

Hyperspectral Imaging for Precision Agriculture

Research · Apr 2026Paper

Wheat yield prediction from hyperspectral data, judged on whether it survives a change of environment.

PythonXGBoostSVRElasticNet1D-CNNRFECV

Problem

Wheat yield decisions get made late and with thin information. Hyperspectral data is enormously wide, heavily correlated, and models trained in one environment tend to collapse in another — which makes single-split accuracy a misleading result.

Approach

  1. 01RFECV-based spectral feature selection to cut a very wide feature space to what carries signal.
  2. 02A model bench spanning SVR, XGBoost, ElasticNet and a 1D-CNN rather than one favoured architecture.
  3. 03A stacked ensemble over those models, evaluated for cross-environment generalization.
How it fits together

Spectral capture

wide, correlated bands

RFECV

cut to what carries signal

Model bench

SVR · XGBoost · ElasticNet · 1D-CNN

Stacked ensemble

combines the bench

Cross-environment eval

the number that counts

Yield prediction

decision support

Role

Framework design, modelling and evaluation.

The trade-off

[DRAFT] Cross-environment evaluation reports worse headline numbers than a single-environment split would. It is also the only number that means anything to a farmer standing in a different field.

Private repo

§06Roadmap

Where this is going

I want to build systems that make good decisions with incomplete information.

It's the thread through everything I've built. A recommendation for an event you have never planned. A match with a stranger you have to trust for the next forty minutes. A yield prediction for a field that hasn't been harvested yet.

The interesting problems were never in the model or the interface. They're in what a system does when it isn't sure — how it handles the case it wasn't trained on, and whether it fails in a way a person can recover from.

That's what I'm building toward: infrastructure that's honest about its own uncertainty, in places where being wrong actually costs someone something.

Currently exploring

  • Agentic automation

    Clawdbot — automating my own workflow before anyone else's.

  • Generative models

    GANs, and what they're actually good for outside demos.

  • GPU compute

    CUDA C — the cost of a computation, not just its correctness.

  • Cloud infrastructure

    AWS — because deployment is where student projects go to die.

  • Precision agriculture

    Taking the hyperspectral work out of the notebook and into a field.

  • End-to-end delivery

    Product to customer to handover, on real client work.

Appendix — the record

Education

  • B.Tech, Computer Science & Engineering

    Bennett University (Times of India Group)

    CGPA 7.79 · 2023 — present

  • Class XII, CBSE

    Greenland Public School, Karnal

    77.8% · 2023

  • Class X, CBSE

    Dyal Singh Public School, Karnal

    90.8% · 2021

Certifications

Achievements

  • INSPIRE Awards, Government of India

    ₹10,000 innovation grant

  • Battle of Brains — 1st place

    Intra-district, MAT & General Knowledge

§07Open a thread

Have an idea worth building?

I'm most useful early — while the problem is still ambiguous and nothing has been decided yet. If that's where you are, start a thread.

Email

mehulg2005@gmail.com

Email me

Elsewhere

GitHub, Mehul Gupta on GitHubLinkedIn, Mehul Gupta on LinkedInRésumé (PDF), one page PDF
Mehul Gupta, standing in a black suit against a plain wall.
Plate 01 — Mehul Gupta2026