RS
CV
Open to SDE / data-focused roles

Rachit Singh —analytics-leddev building acrossweb, ML & APIs.

Full-stack and ML. Shipping clinic systems, training terrain classifiers, building AI apps, often all at once.

Book a call →Download CV
1000+Concurrent users
92%Segmentation accuracy
3Shipped projects
Rachit Singh, developer
Available for new projects
Open

01 — About me

Developer.
Researcher. Builder.

I co-authored a peer-reviewed ML paper on satellite terrain classification, and built Prospera solo, a Gemini-powered career guidance platform used by thousands on launch.

I work across full-stack, ML, and data, not by spreading thin, but because the problems I keep getting pulled toward don't fit a single discipline. Looking for SDE and data-focused roles where that range is useful.

GitHub →LinkedIn →

02 — Featured work

Selected projects.

Three builds. Each deployed, each with a real outcome behind it.

★ Featured92% accuracy — +21 pts2024

Satellite Image Segmentation

PythonTensorFlowU-NetOpenCVStreamlit
Problem

Manual terrain analysis of satellite imagery is slow, expensive, and impractical at scale; existing open-source tools lacked accuracy for remote or low-resolution regions.

Built

Built a deep-learning pipeline using CNN and U-Net architecture with transfer learning and data augmentation. Evaluated 4 model variants and selected U-Net after benchmarking against a baseline of ~71% accuracy. Deployed as an interactive Streamlit demo.

Result

Achieved 92% terrain classification accuracy — a 21-point improvement over baseline. Co-authoring an IEEE research paper on the work. Pipeline processes imagery in seconds versus hours of manual analysis.

Live demo →GitHub →
Satellite Image Segmentation
InterviewForge — AI Avatar Interview Platform
Next.jsFastAPIDocker
Real-time AI interview system2025

InterviewForge — AI Avatar Interview Platform

Problem

Existing interview prep platforms were static and impersonal — questions never adapted to live answers, there was no voice I/O, and candidates had no way to simultaneously practice live coding and verbal reasoning under realistic pressure.

Built

Building an end-to-end AI interview simulation platform: a lifelike 2D/3D avatar (D-ID / HeyGen) conducts real-time mock interviews across DSA, system design, and behavioral rounds. Voice pipeline uses Web Speech API + OpenAI Whisper for STT and ElevenLabs for TTS. GPT-4o / Claude drive adaptive question generation over WebSockets. A Monaco code editor runs inside an isolated Docker sandbox with test-case validation and complexity analysis; an embedded Excalidraw whiteboard handles system-design rounds. Celery + Redis workers handle async audio processing and code execution off the main WebSocket thread.

Result

Full audio → code → AI pipeline running end-to-end in real time. Post-session analytics powered by rubric-based LLM-as-judge scoring. Sandboxed code execution handles arbitrary untrusted input safely — zero escapes in testing.

Live demo →GitHub →
Prospera AI — Career Counseling Platform
Next.jsVercel AI SDKPostgreSQL
1,000+ concurrent users2024

Prospera AI — Career Counseling Platform

Problem

Students had no personalised career guidance tool; generic job boards and one-size-fits-all advice left them uncertain about paths, skill gaps, and opportunities.

Built

Architected an AI career counseling platform with an agentic chatbot and semantic retrieval over pgvector embeddings, served via SSR for fast initial loads. Engineered the full pipeline: Drizzle ORM on PostgreSQL, BullMQ + Redis background jobs, and Auth.js session management with Vercel AI SDK streaming.

Result

Scaled to 1,000+ concurrent users on launch. pgvector-powered semantic search surfaces personalised recommendations in under 200 ms. Open-sourced on GitHub with active forks.

Live demo →GitHub →

03 — Where I've delivered

Experience.

Building production software at fast-moving companies and applying ML to real research.

PhysiNovaLab Pvt Ltd

Software Development Intern — Remote, Bangalore

Feb 2026 — Apr 2026

  • Built responsive UI modules for a full-stack vending machine platform in Next.js 14 (App Router), leveraging SSR and dynamic routing to cut page load times across inventory and transaction views.
  • Integrated Keycloak OAuth 2.0 / OIDC authentication with protected route guards and secure session and token management across the entire application.
  • Consumed REST APIs from a Java Spring Boot backend, surfacing PostgreSQL data models as real-time inventory and transaction management dashboards.

04 — Publications

Research & papers.

Peer-reviewed work at the intersection of deep learning, remote sensing, and environmental monitoring.

Published · 2025

A Deep Learning Expedition Through Satellite Imagery for Environmental Insight

Feature extractionCNN
Global contextTransformer
Temporal dynamicsLSTM

Spatial feature maps from multi-band satellite imagery, extracting edges, textures, and spectral signatures across Landsat and Sentinel data.

95%Train accuracy

Long-range dependency modelling across image patches, capturing spatial relationships that local convolutions miss in complex terrain types.

92%Val. accuracy

Sequential change detection over time-series imagery, identifying environmental shifts in vegetation, water, and urban cover across seasons.

0.20Final train loss

Abstract

Presents a deep learning framework to extract high-resolution environmental insights from multispectral and hyperspectral satellite imagery sourced from Landsat, Sentinel, and commercial satellites. Integrates CNNs for spatial feature extraction, Transformer modules for long-range dependency modelling, and LSTM networks for temporal change detection across four land-cover classes.

Key contributions

  • CNN + Transformer + LSTM fusion for spatial, global, and temporal learning
  • Multi-sensor pipeline: Landsat, Sentinel, SAR, DEM, and meteorological fusion
  • Radiometric correction, cloud masking, augmentation, spectral normalisation
  • 4 land-cover classes: Water, Vegetation, Urban, Barren
95%Train accuracy
92%Val. accuracy
0.20Train loss
0.25Val. loss
20Epochs
4Classes
Megha Sharma1st AuthorRachit Singh2nd AuthorRachit Sharma3rd AuthorAbhishek Verma4th AuthorPragya GaurSupervisor
View full paper →

05 — Expertise

Stack.

What I reach for first.

JavaScript / React / Next.jsTypeScriptTailwind CSSNode.js & REST APIsPython (Pandas, NumPy)TensorFlow / Keras / CNNSQL (MySQL)MongoDBGoogle Analytics APITableau / MatplotlibOpenCVFastAPI / Celery / RedisDockerGit & CI/CDJava
JavaScript / React / Next.jsTypeScriptTailwind CSSNode.js & REST APIsPython (Pandas, NumPy)TensorFlow / Keras / CNNSQL (MySQL)MongoDBGoogle Analytics APITableau / MatplotlibOpenCVFastAPI / Celery / RedisDockerGit & CI/CDJava

06 — Thoughts & writing

Recent articles.

Practical write-ups on React patterns, front-end architecture, and the problems I actually ran into building real projects.

01
TutorialReact · Front-end

Create a Loading Screen in React

Most React apps skip the loading state entirely. Users stare at a blank white screen for half a second before content snaps in. This guide walks through building a polished, animated loading screen using a simple boolean state flag, CSS keyframe animations, and a useEffect cleanup pattern that prevents the dreaded flash on fast connections.

Mar 20245 min read
02
Deep DiveReact · Patterns

Form Validation with Custom Hooks

Repeating validation logic across every form in a codebase is a maintenance nightmare. This article extracts the full validation lifecycle: touched state, error messages, async field checks, and submit locking, into a single reusable useForm hook.

May 20248 min read
03
PatternReact · UI

Pagination Component Patterns

Pagination is deceptively tricky: ellipsis logic, edge-case handling, accessible keyboard navigation, and URL-synced state all need to work together. This breakdown covers three patterns, offset-based, cursor-based, and infinite scroll, with trade-offs for each.

Aug 20247 min read

07 — Verified proof

The work is public.

Certificate, paper, source code. Everything below links to the actual thing.

Published Research

Conference Paper

"A Deep Learning Expedition Through Satellite Imagery for Environmental Insight." Co-authored with 4 peers and a faculty supervisor. 95% training / 92% validation accuracy across 20 epochs, 4 land-cover classes.

5 authors · 2025Read paper
Open Source

Prospera AI GitHub

Full source for the Gemini-powered career guidance platform: publicly auditable code, commit history, and architecture. Built solo in under 6 weeks, handling 1,000+ concurrent users on launch.

Next.js · Gemini API · TailwindView repository
2Companies
1000+Concurrent users
92%Model accuracy
78%ROI improvement
5Research co-authors

08 — Let's build

Let's work
together.

Looking for full-time SDE and ML roles, and open to freelance contracts. I reply same day.

Email me →Schedule a call
Emailrachitdude008@gmail.com
GitHubgithub.com/devxrachit
LinkedInrachit-singh-471491258
LocationIndia, Asia — IST