SA Send Owl

Saksham Agrawal

software engineer · ai/ml

⎇ main available for work IST LeetCode ✓ contest rating

What I do

I build full-stack AI systems, from FastAPI backends and deep-learning models to LLM and RAG pipelines, and turn ideas into working applications.

The Room of Requirement

It gives you exactly what you need. Here: five things I built.

Every project has its own Daily Prophet front page. Open one to read the full story.

Late EditionProject · Live

The Daily Prophet

Vol. CCXXVI · No. TodayPrice 5 Knuts

JobGenie Conjures a Smarter Way to Apply

Resume intelligence, ATS simulation and grounded AI analysis, so every application is better informed.

Job searching involves more than finding openings. Candidates also need to understand how their experience matches each role. JobGenie brings resume parsing, ATS-score simulation and job-description matching into one application copilot, helping candidates make better decisions before they apply.

At its core, JobGenie uses a RAG pipeline to ground LLM outputs and reduce hallucination in match scoring. An LLM analytics dashboard shows how each application fits, while multi-resume A/B testing makes it easy to compare versions of a resume.

The platform also tracks interviews to keep the whole process organised. The backend runs on Python and FastAPI, with PostgreSQL for data and Docker for shipping. It is live at jobgenie.tech.

Resume · Job description
FastAPI backend
Gemini API
RAG grounding
Match score · ATS simulation
A resume and a job description flow through FastAPI and RAG-grounded Gemini analysis to a match score.

Late EditionProject · 5th Place

The Daily Prophet

Vol. CCXXVI · No. TodayPrice 5 Knuts

PhoniX Unmasks the Voice Behind the Deception

A Flutter app combines live call audio analysis with deep-learning detection to flag AI-generated voices.

Voice phishing is harder to recognise as AI-generated voices become more convincing. PhoniX tackles this by analysing incoming call audio in real time. Its pre-trained TensorFlow model examines voice characteristics and alerts the user when a voice is flagged as deepfake or AI-generated.

When a call comes in, the app intercepts its audio stream and feeds it into the detection model. The model analyses the voice and raises real-time alerts, giving users an extra layer of awareness during a potentially fraudulent call.

PhoniX is built as a Flutter mobile app, with Firebase for secure sign-in and data storage. The model reached 89.68% accuracy, and the project took fifth place at CSA Hackathon 3.0 at UPES.

Incoming call audio
intercepted live
TensorFlow model
voice characteristics
Human or AI-generated
Real-time alert
Incoming call audio flows through TensorFlow detection to a real-time alert.

Late EditionProject · Research

The Daily Prophet

Vol. CCXXVI · No. TodayPrice 5 Knuts

Three Architectures, One Recommendation Challenge

An experimental comparison of recurrent networks and Kolmogorov-Arnold Networks on real e-commerce clickstream data.

What makes a good next-item recommendation? This research project explores three neural architectures for session-based recommendation on the YooChoose e-commerce clickstream dataset: a GRU-based recurrent model, a custom KAN and a hybrid of the two.

The KAN was implemented from scratch with B-spline layers and attention pooling. The hybrid lets a KAN layer refine the GRU output. All three models share one training pipeline, so the comparison stays fair.

The experiments use 50K sessions and more than 128K samples, benchmarked on HR@10, NDCG@10 and accuracy. The best results reached 58.08% HR@10 and 39.42% NDCG@10.

YooChoose sessions
50K sessions · 128K+ samples
GRU
Recurrent KAN · GRU + KAN
Next-item ranking
HR@10 58.08%
NDCG@10 39.42%
YooChoose sessions feed three architectures through one shared training and evaluation pipeline.

Late EditionProject · Open Source

The Daily Prophet

Vol. CCXXVI · No. TodayPrice 5 Knuts

Launch Cadence: Mission Control, Simplified

A role-based rocket launch management system built with nothing but Python’s standard library and flat files.

Launch Cadence brings rocket launch management and customer bookings into a single console application. It separates admin work from customer workflows: admins manage rocket specifications and launch schedules, while users browse rockets and reserve launch slots.

Admins set rocket dimensions, mass, payload capacity to LEO, GTO and SSO, and pricing. They also schedule launch slots with leap-year-aware date validation. Users can sign up, log in, browse what is available and book a launch.

The app keeps active bookings, booking history and per-user remarks. Everything is stored in plain text files and it uses only the Python standard library, so it runs anywhere with no dependencies.

Admin
rockets · pricing · launch slots
Plain .txt files
User
browse · book · history
Admin-managed rockets and schedules connect to customer bookings through flat-file storage.

Late EditionProject · Team Project

The Daily Prophet

Vol. CCXXVI · No. TodayPrice 5 Knuts

The Library Ledger: Every Book Accounted For

A desktop library management system that connects book and member records to a relational database.

Running a library means keeping track of both the collection and its members. This team project is a Java Swing desktop application for managing books and member records, with database-backed workflows for issuing and returning books.

The app brings core library operations into one desktop interface. Librarians work with book and member records through Swing screens, while issue and return workflows handle circulation.

A relational SQL database stores every record. The project pairs Java desktop development with database operations for a structured approach to running a library.

Java Swing GUI
Books · Members · Issue / Return
Relational database
The Swing interface connects book and member operations to a relational SQL database.

Put on the Sorting Hat

“Hmm… a new head. Shall I sort you, or tell you about Saksham?”

The Wizarding Journey

  1. Intern · IBM

    • Trained a custom CNN in PyTorch on the PlantVillage image dataset across 15 disease classes and benchmarked it against EfficientNet, shipping the custom model after it scored ~10% higher on macro-F1.
    • Deployed the classifier as a Streamlit application that diagnoses leaf images and returns treatment guidance in 2 languages (Hindi and English), combining stored recommendations with LLM-generated output.

    Python · PyTorch · CNN · Computer Vision · Streamlit

    View on GitHub →
  2. President · Cloud Security Alliance (CSA), UPES

    • Lead the UPES CSA chapter and its 50-member team, overseeing technical initiatives, industry partnerships, and execution of community events.
  3. Registrations Head · Cloud Security Alliance (CSA), UPES

    • Led the registrations team and ran end-to-end registration and onboarding operations for AWS Community Day Dehradun, supporting 400+ participants across cross-functional event teams.
  4. 5th Place · CSA Hackathon 3.0, UPES

    • Recognised for innovation in AI-driven cybersecurity solutions (PhoniX project).
  5. Social Intern · Gram Rajya Vikas Evam Prashikshan Sansthan

    • Assisted legal authorities in more than 5 rescue and rehabilitation operations related to child-labour cases.
    • Supported more than 20 interventions under child protection and POCSO regulations.

    Field Coordination · Documentation · Community Engagement · Teamwork

The Spellbook

The languages, tools and techniques I use to turn ideas into working software.

Gold spells are ones I have used in a real project.

II

AI / ML

Models designed, trained and benchmarked from scratch, and LLMs grounded with RAG so their answers hold up.

  • PyTorch (used in a project)
  • TensorFlow (used in a project)
  • CNN (used in a project)
  • GRU / LSTM (used in a project)
  • KAN (used in a project)
  • Gemini API (used in a project)
  • RAG (used in a project)
  • scikit-learn
  • OpenCV
  • NumPy
  • pandas
  • Vector Search

Used in JobGenie, Session recommender, PhoniX, IBM internship (custom CNN that beat EfficientNet on macro-F1)

III

Backend & Data

APIs and data stores built with production in mind.

  • FastAPI (used in a project)
  • PostgreSQL (used in a project)
  • REST APIs
  • MySQL
  • MongoDB
  • Redis

Used in JobGenie

IV

Tools

Shipping, packaging and the interfaces people actually click.

  • Docker (used in a project)
  • Streamlit (used in a project)
  • Swing (used in a project)
  • Git
  • CI/CD

Used in JobGenie, IBM internship (Streamlit app in Hindi and English), Library system

New · Spell Class Practise real commands as spells Shell, git and Docker, taught by a professor. Win points for your house. Enter the classroom →

Hogwarts Records

Saksham Agrawal in Hogwarts robes
Saksham AgrawalPoint the wand at the wall
I don’t like leaving ideas as ideas. If I can imagine it, I want to build it.

I’m a final-year Computer Science (AI & ML) student at UPES who enjoys turning concepts into real products. From training machine learning models and building FastAPI backends to integrating LLMs with RAG, I love working across the entire stack. Whether it’s JobGenie or my latest experiment, I’m always chasing the next problem worth solving.

Outside of building, you’ll usually find me solving DSA, exploring new technologies, or figuring out how to make software just a little smarter.

UPES · Office of the Registrar

B.Tech, Computer Science & Engineering (AI/ML)

· CGPA 8.75 / 10

Relevant coursework: Machine Learning · Pattern & Visual Recognition · Deep Learning · Data Structures & Algorithms · Database Management Systems · Operating Systems.

The House Points Ledger

Opening the ledger…

The Owl Post

Have something that needs to be built? Open to freelance and full-time opportunities. Reach out about a role, a project, or just to talk shop. I read everything myself.

Send an owl →

Write a Howler

Say what you need. It arrives straight in my inbox, and I reply myself.

↑↓ move · ↵ cast · esc close