NEXUS B.Tech CSE (AI/ML) student

Aasveer Singh

AI/ML Engineer

RAG systems, LLM pipelines, NLP, Recommendation engines

Building end-to-end AI applications with Python, SQL and FastAPI: multi-stage LLM pipelines, retrieval-augmented generation and NLP, served through REST APIs.

Focus

Generative AI
Large Language Models
RAG
NLP
Machine Learning
Recommendation Systems

Amritsar, Punjab, India

Download resume

GitHub

Projects: 6+
Certifications: 12+
Verified skill links: 45+

The engineer behind the systems

I build practical AI systems, from intelligent pipelines to experiences people can actually use.

Aasveer Singh

AI/ML Engineer

Education: B.Tech, Computer Science & Engineering (AI/ML)
Amritsar Group of Colleges, 2023–2027
Based in: Amritsar, Punjab, India

I'm Aasveer, a Computer Science & Engineering (AI/ML) student who enjoys building practical applications that bring intelligent systems, backend engineering, data processing and usable interfaces together.

Drawn to

Generative AI
LLMs
RAG
NLP
Machine Learning
Recommendation Systems
AI-powered applications

Engineering perspective

Problem

Start with what a person is actually trying to do, and what a useful result would look like, before choosing any tools.

Grounded in

AI-powered applications

A way of looking at a problem, not the architecture of any one project.

Problem

Start with what a person is actually trying to do, and what a useful result would look like, before choosing any tools.

Intelligence

Choose the technique that fits the problem, whether that is retrieval, language understanding, ranking or a language model.

Engineering

Build what surrounds the model: the data processing, backend services and APIs that let it run as a real application.

Experience

Bring it to an interface someone can use, because that is where all of the other work becomes useful.

Engineering philosophy

I enjoy building practical applications that use AI to address real problems rather than treating models as isolated demonstrations.
I am interested in bringing together intelligent systems, data processing, backend engineering, APIs and user-facing interfaces.
The usefulness of an intelligent system depends on how well its different components work together for the person using it.

How I build

A way of thinking more than a fixed recipe. Every project moves through these differently.

Understand the problem

Work out what the person actually needs and what a good result looks like, before choosing any tools.

Seen in GiftBro AI

The real risk was a model inventing gifts that don't exist, so ranking real catalogue items became the core of the design.

Understand the problem

Work out what the person actually needs and what a good result looks like, before choosing any tools.

Design the system

Decide what the parts are, what each one is responsible for, and how data moves between them.

Build the pipeline

Build the main path one stage at a time, so each step can be checked on its own.

Connect the components

Join the backend, data and interface through clear boundaries, so one part can change without breaking the rest.

Test and refine

Try real inputs, see what breaks, and plan for what happens when a dependency isn't there.

Engineering connections

Where my skills meet real builds.

An overview of the technical capabilities applied across my projects.

16 skills · 6 projects · 45 verified links

View the connections as a table

Skills and the projects that use them
Skill	Used in
RAG	Wingman AI: Local knowledge base searched with TF-IDF and cosine similarity. StudyForge AI: Source-grounded answers from retrieved chunks of uploaded PDFs. InterviewIQ: Knowledge base that stores notes as paragraph chunks and ranks them by keyword matches.
NLP	Wingman AI: Signal analysis of conversations: sentiment, reciprocity, questions and energy. StudyForge AI: Text processing for retrieval over study material, within the approved contribution.
OCR	Wingman AI: Reads chat screenshots with Tesseract through pytesseract.
TF-IDF	Wingman AI: Vectorizes knowledge-base entries and the conversation for retrieval. GiftBro AI: Vectorizes catalogue items and the recipient profile. StudyForge AI: Vectorizes PDF chunks and the student's question.
Cosine Similarity	Wingman AI: Ranks knowledge-base entries against the conversation. GiftBro AI: Provides the semantic signal in the weighted ranking. StudyForge AI: Ranks PDF chunks against the question.
Recommendation Systems	GiftBro AI: Hybrid recommender that ranks real catalogue items.
Scikit-learn	Wingman AI: TfidfVectorizer and cosine_similarity for retrieval. GiftBro AI: TfidfVectorizer and cosine_similarity for scoring. StudyForge AI: TfidfVectorizer and cosine_similarity for retrieval.
Gemini API	InterviewIQ: Generates questions and evaluates answers, resumes and code. GiftBro AI: Optional personalization layer around the ranked shortlist. StudyForge AI: Writes answers, flashcards and quizzes from study material.
FastAPI	InterviewIQ: Backend API with separate routers for each feature. Wingman AI: REST API for text and screenshot analysis. GiftBro AI: REST API exposing the recommendation pipeline. StudyForge AI: Backend for accounts, documents, chat and study tools. PennyBloom: A thin API layer over the shared Python core, with no business logic.
SQLite	InterviewIQ: Users, interviews, evaluations and documents, through SQLAlchemy. Wingman AI: Saved items and analysis history. GiftBro AI: Recommendation history and favorites. StudyForge AI: Accounts, chunks, quizzes, tasks and conversations.
Streamlit	Wingman AI: The multi-page application interface. GiftBro AI: The recommendation interface.
JWT Authentication	InterviewIQ: Login returns a signed JWT used as a bearer token.
React	InterviewIQ: The React 19 application views. StudyForge AI: The React interface, built with Next.js. SmartQueue AI: The whole application, including shared queue state through Context. PennyBloom: The web interface: dashboard, transactions, reports and budgets.
TypeScript	InterviewIQ: The typed frontend and the Express server. StudyForge AI: The typed Next.js frontend. SmartQueue AI: Typed tokens, counters and the rule engine. PennyBloom: The typed web app and its API client.
Vite	InterviewIQ: Builds the app and serves it through the Express server. SmartQueue AI: The build tool and development server. PennyBloom: The build tool and development server.
Tailwind CSS	InterviewIQ: Styles the application views. SmartQueue AI: Styles the application views. PennyBloom: Styles the web interface.

Don't take my word for it.

Every technical claim should have something behind it. Explore the projects, implementations and engineering evidence that demonstrate how I turn technical knowledge into working systems.

10 capabilities · 20 implementations · 41 inspected source files

Evidence record 01 / 10

Retrieval-Augmented Generation

Answers grounded in retrieved source text, not in a model's memory.

CapabilityRetrieval-Augmented Generation
ProjectStudyForge AI
ImplementationPDF chunking, TF-IDF retrieval and source-grounded answers
Evidence4 inspected files, 4 public

Engineering context

Uploaded PDFs are split into overlapping, page-numbered chunks. A question retrieves up to four chunks by TF-IDF and cosine similarity, and Gemini answers from them. Without an API key, the retrieved text is shown and labelled instead.

My contributionConfirmed contribution

RAG, study planning, quizzes, NLP, FastAPI and SQLite.

Engineering insight

Designs for the failure case: when the model is unavailable the feature still works, and says plainly what it is showing.

Retrieval-Augmented Generation

StudyForge AI: Uploaded PDFs are split into overlapping, page-numbered chunks. A question retrieves up to four chunks by TF-IDF and cosine similarity, and Gemini answers from them. Without an API key, the retrieved text is shown and labelled instead.

Wingman AI: Hand-written guidance entries are vectorized with TF-IDF over words and word pairs, and the three closest to the conversation are returned with their scores.

Generative AI & LLM Applications

StudyForge AI: Gemini reads the PDF and returns flashcards or quiz questions as JSON. Each item is checked field by field, repeats are dropped, and when the provider is busy a quiz is built from the document's own sentences.

GiftBro AI: A deterministic recommender ranks real catalogue items first. Gemini then explains the best fit and suggests a message, constrained by a response schema and told to use only the supplied candidates.

InterviewIQ: Gemini generates role-specific questions, then a second call scores every answer and four separate dimensions. Question counts, option sets and every score range are validated before use.

Natural Language Processing

Wingman AI: A conversation is tokenized and scored separately for sentiment, questions, reciprocity and energy. Fixed weights combine the signals into a next move and reply strategies.

StudyForge AI: Study questions are reduced to meaningful words and compared by overlap, so generated material does not ask the same idea twice.

backend/app/main.py

OCR & Document Intelligence

Wingman AI: Chat screenshots are read with Tesseract OCR, then passed into the same signal analysis and retrieval as typed text, through a dedicated API endpoint.

Recommendation Systems

GiftBro AI: Gifts over budget or on the avoid list are removed. The rest are scored with TF-IDF and cosine similarity against interests plus separate match scores, combined with fixed weights into one ranking.

Backend & API Engineering

StudyForge AI: One FastAPI service handles accounts, PDF uploads, chat, flashcards, quizzes and planning, with an automated test suite for its API.

Wingman AI: The same analysis is exposed over HTTP for text and for screenshots, separately from the Streamlit interface.

api.py

PennyBloom: The web API calls the same finance core as the command-line app, and holds no business logic of its own, so the two interfaces cannot disagree.

InterviewIQ: Interviews, analytics, the knowledge base and authentication are separate routers mounted on one application.

backend/app/main.py

Database Integration

StudyForge AI: Documents, page-numbered chunks, flashcards, quizzes and quiz attempts are stored in SQLite and queried directly.

backend/app/main.py

Wingman AI: Analyses and saved replies are kept in SQLite for later review.

core/db.py

InterviewIQ: Users, interviews, evaluations and knowledge-base documents are SQLAlchemy models. Local development uses SQLite, and the Docker setup runs PostgreSQL.

Frontend & Backend Integration

InterviewIQ: A React app is served by Express and Vite, which forward API calls to the FastAPI backend, so the browser talks to a single origin.

server.ts

PennyBloom: The React and TypeScript web app reads and writes through one typed API client for the FastAPI layer.

frontend/src/api.ts

Data Processing & Integrity

PennyBloom: Transactions become monthly summaries, category breakdowns, a least-squares trend and a stated next-month estimate. Files are written atomically and backups carry SHA-256 checksums.

Queue Management Logic

SmartQueue AI: Visitors join and receive numbered tokens. Waiting time and backlog are estimated from token timestamps and open counters, and staff call the next token to a specific counter.

6 case files

Featured projects

Case file 01 / 06
01
StudyForge AI
AI-Powered Study Platform for CS Students
CS students studying from their own PDFs need answers grounded in that source material, along with practice material and tools to plan and track revision.
- Python
- FastAPI
- Next.js
- React
- +3
Case file 02 / 06
02
PennyBloom
Personal Finance Management Platform
Managing personal money means recording income and expenses, watching monthly budgets and remembering recurring payments, with records that stay safe and easy to understand.
- Python
- FastAPI
- React
- TypeScript
- +4
Case file 03 / 06
03
InterviewIQ
AI-Powered Mock Interview Platform
Students preparing for technical, behavioral (STAR) and system-design interviews need realistic practice questions and structured feedback on each answer.
- React 19
- TypeScript
- Tailwind CSS
- Python
- +7
Case file 04 / 06
04
Wingman AI
AI-Powered Conversation Intelligence Platform (RAG + NLP)
Chat screenshots hold interaction signals such as warmth, reciprocity and engagement that are hard to turn into clear next-step recommendations, reply strategies and date plans.
- Python
- FastAPI
- Streamlit
- OCR
- +6
Case file 05 / 06
05
GiftBro AI
Hybrid ML + Generative AI Gift Recommendation System
Gift suggestions need to fit the recipient's relationship, occasion, interests, personality, budget and things to avoid, and an AI model should never invent products that do not exist.
- Python
- Scikit-learn
- Pandas
- TF-IDF
- +5
Case file 06 / 06
06
SmartQueue AI
Intelligent Queue Management & Analytics System
Physical waiting lines give visitors no view of their position or likely wait, and give staff no single place to dispatch people across several service counters or to see where the queue is backing up.
- React
- TypeScript
- Vite
- Tailwind CSS
- +2

8 categories

Technical skills

Languages
- Python
- SQL
- C/C++
- JavaScript
AI / ML
- Machine Learning
- Feature Engineering
- Forecasting (Prophet)
Generative AI & NLP
- LLMs
- Generative AI
- Prompt Engineering
- Information Retrieval
- AI Output Evaluation
AI Models & Platforms
- OpenAI API
- Anthropic Claude API
- Claude Code
- Hybrid ML + LLM pipelines
Backend & APIs
- Flask
- REST APIs
- SQLAlchemy
- Python-dotenv
Data & Databases
- Pandas
- NumPy
- Matplotlib
- Data Analysis
- PDF Parsing (PyPDF2)
- MySQL
- PostgreSQL
Frontend
- HTML
- CSS
- React Context API
Tools & Concepts
- Git
- GitHub
- Jupyter Notebook
- Google Colab
- OOP
- DBMS
- Data Structures
- Problem Solving

3 case files

Behind the build.

Every system has an engineering story. Explore the challenges, investigations and technical decisions behind the systems I build.

Case 01Retrieval and reliability
StudyForge AI
Answers that stay with the source
A study assistant that answers from a student's own PDFs, and keeps working, honestly labelled, when the language model cannot.
Case 02OCR, NLP and local retrieval
Wingman AI
From a screenshot to a decision brief
Turning an image of a conversation into measured signals and written guidance, each of which can be explained.
Case 03Recommendation and generative AI
GiftBro AI
Ranking first, generating second
A recommender in which the language model can explain a choice, but never invent one.

Case 01StudyForge AIRetrieval and reliability

Answers that stay with the source

The challenge
Students ask questions about the PDFs they upload. An answer is only useful if it comes from that material and can be traced back to a page. The assistant also depends on an external model provider, which may not be configured or may be busy.
The investigation
The code distinguishes three situations that each need a different, honest response: a model is available and answers, no model is configured, and the provider is busy. Retrieval therefore has to work on its own, without the model, and every answer has to say which of these it came from.
03
The approach
- PDF text is read page by page and split into 180-word chunks that overlap by 30 words, each stored in SQLite with its page number.
- A question is matched against the chunks with TF-IDF and cosine similarity, and up to four chunks with a score above zero are kept.
- When the model answers from the PDF, the page citations in its answer are read back and returned with that page's text as the sources.
- Without a configured model, the retrieved passages themselves are returned in a labelled source-only mode, with a message saying the model is not configured.
- Generated flashcards and quizzes are checked field by field, near-duplicate questions are dropped by comparing their meaningful words, and when the provider is busy a quiz is built from the document's own sentences.
04
The result
- Questions about an uploaded PDF are answered with page-level sources, and the feature still returns the relevant passages when no model is available.
- The backend has tests for these paths, including test_source_only_mode, test_provider_failure_is_retryable, test_local_quiz_is_available_when_provider_is_busy, test_gemini_incomplete_finish_is_not_success and test_study_items_keep_distinct_concepts_and_add_focus.
The takeaway
Design for the failure case. When the model is unavailable the feature should still work, and say plainly what it is showing.

Case 02Wingman AIOCR, NLP and local retrieval

From a screenshot to a decision brief

The challenge
Conversations usually arrive as screenshots, not text. The analysis has to start from an image, and its conclusions need to be explainable rather than a single opaque score.
The investigation
The README sets the principle: "Read patterns, not single messages." That rules out judging one message, or one number, and points to separate signals measured across the whole conversation, which must behave the same whether the text was typed or read from an image.
03
The approach
- Tesseract OCR reads the screenshot. Its location comes from TESSERACT_CMD or the standard Windows install paths, and failures return a readable message, such as no text detected or Tesseract not found, instead of an exception.
- The extracted text goes through the same analysis as typed text: sentiment, questions, reciprocity and energy are scored separately, and fixed weights combine them (sentiment 0.35, reciprocity 0.30, flirt 0.20, questions 0.15).
- Guidance comes from a local knowledge base searched with TF-IDF over words and word pairs, and the three closest entries are returned with their scores.
- A FastAPI endpoint accepts a screenshot, rejects anything that is not an image with a 400, returns a 422 when no text can be read, and otherwise returns the extracted text with the full analysis.
04
The result
- One analysis core serves both the Streamlit app and a REST API, for typed conversations and for screenshots.
- Every recommendation can be traced to measured signals and to named entries in the knowledge base.
The takeaway
Connect OCR to a downstream pipeline instead of stopping at extracted text, and prefer transparent signals so that every result can be explained.

Case 03GiftBro AIRecommendation and generative AI

Ranking first, generating second

The challenge
Asking a language model for gift ideas directly can return items that do not exist, ignore the budget, or cannot be compared. The README states the alternative: "Instead of simply asking an AI to invent gift ideas, GiftBro first ranks real catalogue items."
The investigation
A recipient profile carries many separate signals: interests, occasion, relationship, personality, style, budget and things to avoid. Some are constraints and some are preferences. That points to filtering on the constraints first, scoring each preference separately, and keeping the model out of the ranking.
03
The approach
- Items on the avoid list are removed, but only while some candidates remain. Items within budget form the pool, and if nothing fits, the cheapest items are used instead, so the list is never empty.
- TF-IDF over words and word pairs, with cosine similarity, compares the profile with each item, and separate scores measure interest, occasion, relationship, personality, style, personalization and budget fit.
- Fixed weights combine the scores into one ranking (semantic 0.20, interest 0.20, occasion 0.15, relationship 0.12, personality 0.10, budget 0.10, style 0.08, personalization 0.05), and the shortlist favours different categories before repeating one.
- Gemini runs only when the user asks. It receives the ranked candidates, is told to recommend only those, and must answer in a fixed response schema. If it is unavailable or fails, nothing is returned and the ranking stands.
04
The result
- A ranked shortlist of real catalogue items, with every score readable, exists before any text is generated.
- Generation is optional and limited to explaining, personalizing and presenting candidates the recommender has already chosen.
The takeaway
Keep the decision with a deterministic system, and give the language model a narrow job whose output can be checked.

7 systems

The AI Systems Laboratory

Capability map

Where each capability is built.

Select a capability to see the projects that use it, or a project to see what it is built from. Every line is checked against that project's source.

Capability

RAG and Retrieval Systems

Finding the right source text first, so a result rests on real material.

Connected projects 3

Wingman AI
TF-IDF retrieval over a local knowledge base.
StudyForge AI
Retrieval over chunks of uploaded PDFs.
InterviewIQ
A knowledge base ranked by keyword matches.

RAG and Retrieval Systems

Finding the right source text first, so a result rests on real material. Used in: Wingman AI (TF-IDF retrieval over a local knowledge base.); StudyForge AI (Retrieval over chunks of uploaded PDFs.); InterviewIQ (A knowledge base ranked by keyword matches.).

Generative AI and LLM Applications

Using language models where they help, inside a system that keeps them on track. Used in: InterviewIQ (Gemini generates questions and evaluates answers.); GiftBro AI (Gemini personalizes a shortlist it cannot change.); StudyForge AI (Gemini writes answers, flashcards and quizzes.).

NLP

Turning human language into signals and scores that software can act on. Used in: Wingman AI (Signal analysis of conversations.); StudyForge AI (Text processing for retrieval over study material.).

Recommendation Engines

Ranking real items against a profile, in a way that can be explained. Used in: GiftBro AI (Hybrid TF-IDF and weighted ranking of catalogue gifts.).

Backend and API Engineering

The services, data and contracts that make everything else dependable. Used in: InterviewIQ (FastAPI routers with JWT authentication.); Wingman AI (A REST API over the analysis core.); GiftBro AI (A REST API over the recommender.); StudyForge AI (FastAPI for accounts, documents and chat.); PennyBloom (A thin FastAPI layer with no business logic.).

Intelligent Data Applications

Turning everyday records into something a person can understand and use. Used in: SmartQueue AI (Rule-based congestion insights from queue state.); PennyBloom (Reports, budgets and trends from personal records.).

Machine Learning

Learning from data. Here it vectorizes text and scores candidates; none of these projects trains a predictive model. Used in: Wingman AI (Scikit-learn TF-IDF vectors for the knowledge-base search.); StudyForge AI (Scikit-learn TF-IDF vectors for retrieval over PDF chunks.); GiftBro AI (Scikit-learn similarity plus engineered match scores for each gift.).

OCR and Document Intelligence

Reading the text inside images and documents, so it can be analysed. Used in: Wingman AI (Reads chat screenshots with Tesseract OCR.); StudyForge AI (Extracts PDF text page by page with PyMuPDF, then chunks it.).

AI Models and Platforms

The model APIs behind the generative features, and calling them reliably. All three use Google's Gemini API. Used in: InterviewIQ (Gemini through the GenAI SDK, with three attempts per request.); StudyForge AI (Gemini with a fallback model, and Groq for streamed chat.); GiftBro AI (Gemini with a response schema it must follow.).

Data and Database Systems

Where the data lives, and how it is queried. Used in: StudyForge AI (SQLite with hand-written SQL for documents, chunks and study material.); Wingman AI (SQLite for saved analyses and history.); GiftBro AI (SQLite for recommendation history.); InterviewIQ (SQLAlchemy over SQLite locally; its Docker setup runs PostgreSQL.).

Frontend Engineering

The interfaces people actually use, from React apps to Python dashboards. Used in: StudyForge AI (A Next.js and React study workspace.); PennyBloom (A React, TypeScript and Tailwind CSS finance app.); InterviewIQ (A React, TypeScript and Tailwind CSS interview platform.); SmartQueue AI (React, TypeScript and Tailwind CSS queue screens.); Wingman AI (A Streamlit interface for the analysis.); GiftBro AI (A Streamlit interface for recommendations.).

Analytics and Forecasting

Turning records into figures, trends and stated estimates. Used in: PennyBloom (Reports, a least-squares spending trend and a next-month estimate from a three-month average.); SmartQueue AI (Wait and service times, priority share and no-shows from queue tokens.); InterviewIQ (Score history and topic averages across interviews.).

System workflows

Open a system and follow its pipeline stage by stage. Every stage is taken from the project's own code, including the stages a project does not have.

Retrieval-Augmented Generation

Retrieval finds the most relevant text before an answer is written, so the result rests on real source material. Wingman AI and StudyForge AI both retrieve with TF-IDF and cosine similarity, but differ in what they index and in what happens after retrieval. InterviewIQ is simpler again: it ranks paragraphs by keyword matches, with no vectors and no language model. Switch between them to compare.

Ingest: Prepares source text so it can be searched.
Retrieve: Scores the source text against the question.
Respond: Turns the retrieved text into what the user sees.

Stage 01 of 07 / Ingest

Document input

Every retrieval system starts with a defined body of source text. Answers can only be grounded in what was put here.

In Wingman AI

A curated local knowledge base. Each entry in knowledge_base.json is hand-written guidance with a title, category, tags and content.

Input: knowledge_base.json, read from disk at retrieval time
Output: A list of guidance entries

core/rag.py

KB=Path(__file__).resolve().parents[1]/'knowledge_base.json'

Education

B.Tech, Computer Science & Engineering (AI/ML), Amritsar Group of Colleges, 2023–2027

Earlier education

2023Class XIIShri Guru Harkishan Sen. Sec. School
2021Class XShri Guru Harkishan Public School

Credential orbit // Spatial credential environment

Certifications

View every credential as a list

Claude Code 101Anthropic Claude Academy
Building with the Claude APIAnthropic Claude Academy
Introduction to Large Language ModelsIBM SkillsBuild
Machine LearningIBM SkillsBuild
Introduction to Generative AIGoogle Cloud Skills Boost
Introduction to Modern AICisco
Introduction to Prompt Engineering with GitHubMicrosoft Learn
Azure Database for PostgreSQL: Vector SearchMicrosoft Learn
Claude Platform 101Anthropic Claude Academy
Claude 101Anthropic Claude Academy
AI Practitioner: Ready to Use AIIBM SkillsBuild
Critical Thinking in the AI EraHP LIFE

Let's start a conversation.

Have a project in mind, an opportunity to discuss or a technical question? Send me a message.

Aasveer Singh

I build practical AI systems, from intelligent pipelines to experiences people can actually use.

AEngineering perspective

Problem

BEngineering philosophy

CHow I build

Understand the problem

Where my skills meet real builds.

Retrieval-Augmented Generation

Engineering context

My contributionConfirmed contribution

Engineering insight

StudyForge AI

PennyBloom

InterviewIQ

Wingman AI

GiftBro AI

SmartQueue AI

Languages

AI / ML

Generative AI & NLP

AI Models & Platforms

Backend & APIs

Data & Databases

Frontend

Tools & Concepts

Answers that stay with the source

From a screenshot to a decision brief

Ranking first, generating second

Where each capability is built.

RAG and Retrieval Systems

Retrieval-Augmented Generation

Document input

B.Tech, Computer Science & Engineering (AI/ML), Amritsar Group of Colleges, 2023–2027

Earlier education

Certifications

Engineering perspective

Engineering philosophy

How I build