Ankit Paul

I’m a Computer Science undergraduate working in machine learning and scalable AI systems. I have interned at Vaani as a Machine Learning Engineering Intern, developing load-aware LLM gateways and citation-grounded RAG research pipelines. I’ve also interned at FlyRank as a Machine Learning Intern, building search ranking decay prediction models and automated data contract auditing pipelines.

My technical work spans AI engineering (retrieval-augmented generation, LLM admission control gateways, multimodal agent systems), ML engineering (deep learning representation, ranking algorithms, automated data contracts), end-to-end software architecture & full-stack systems, and quantitative finance (stochastic modeling, backtesting, statistical analysis).

I enjoy building and deploying products, developer tools, and autonomous software agents, taking ideas from concept to real-world systems. Outside of engineering, you’ll usually find me playing football, capturing moments through photography and videography, sketching & drawing, or exploring open-source models.

Ankit Paul
please hire me !
ankitpaul9341@gmail.com GitHub LinkedIn Email Instagram

Achievements

Experience

Vaani Logo
Machine Learning Engineering Intern, Vaani
LLM Gateway Architecture & Grounded RAG Systems
Nov 2026
More Info
Centralized Choke-Point LLM Gateway
  • Accomplished global rate limit compliance and zero upstream 429 errors across 300+ requests, as measured by sustaining burst loads up to 61.5 requests/min peak offered rate without queue collapse, by designing a centralized choke-point Gateway featuring dual token buckets, concurrency semaphores, and Earliest Deadline First (EDF) priority scheduling.
  • Accomplished sub-millisecond deadline shedding and cost containment during provider exhaustion, as measured by reducing marginal token expenditure to 0 during quota limits, by engineering an L0–L4 adaptive degradation ladder that automatically sheds expired requests and smoothly falls back to local zero-token extractive search.
  • Accomplished multi-tenant system resilience, as measured by eliminating client queue stampedes and tracking 200,000 token/day ceilings, by implementing concurrency semaphores, rolling retry budgets with exponential backoff, and persistent daily quota state stores.
Zero-Hallucination Grounded RAG Pipeline
  • Accomplished zero-hallucination question answering across 3 arXiv machine learning papers, as measured by 100% citation entailment coverage or truthful refusal, by constructing an end-to-end grounded RAG pipeline with post-generation attribution validation.
  • Accomplished strict claim traceability, as measured by verifying fine-grained paragraph-level and section-level attribution for every generated claim ([{paper_id}v{version}, p.{page}, §{section}]), by implementing a two-phase semantic entailment and citation checking stage.
  • Accomplished benchmark reproducibility across synthetic burst and sustained overload tests, as measured by 0 rate-limited responses and 100% query survival, by coupling the RAG pipeline directly to the centralized admission control gateway.
FlyRank Logo
Machine Learning Intern, FlyRank
SEO Search Ranking Prediction & Representation Learning
Aug 2026 - Oct 2026
More Info
Transformer Ranking Models & Feature Engineering
  • Accomplished search ranking position prediction on 30,000+ historical search-log rows, as measured by achieving 84% Precision@50 and 91% Top-20 Precision (a 3.5x lift over the 24% rule-based baseline), by training Random Forest and transformer ranking models.
  • Accomplished audit repeatability and reduced setup overhead across 5 client benchmark runs, as measured by saving 2.0 hours per benchmark run, by building a 4-step automated data contract validation pipeline.
  • Accomplished leak-free model generalizability on out-of-sample client domains, as measured by maintaining 0.00% target feature leakage across 22,006 active content pages, by enforcing strict data contract boundaries and GroupKFold cross-validation.
  • Accomplished model transparency and domain-drift diagnosis, as measured by authoring a comprehensive 38-page engineering evaluation report, by analyzing model drift, inference latency bottlenecks, and Precision-Recall trade-offs across search verticals.
Google Logo
Google Student Ambassador
Campus Community & Developer Programs
2024 - Present
More Info
  • Accomplished developer ecosystem engagement and technical capability building, as measured by mentoring 100+ students and generating 8,000+ developer impressions across 20+ technical workshops, by delivering hands-on architecture walkthroughs and live coding labs with Google Cloud and Gemini AI APIs.
  • Accomplished developer skill acceleration among student engineers, as measured by organizing multiple study jams and hackathons, by creating structured code walkthroughs, live architecture demos, and algorithmic problem-solving guides.

Projects

The Relay: Resilient LLM Gateway & Grounded RAG System
Python, Groq Cloud API, Stand-In Server, BM25 Retriever, EDF Scheduling, Dual Token Buckets
Nov 2026 More Info
  • Internship Project: Built during my Machine Learning Engineering Internship at Vaani as the core capstone research and LLM inference engine.
  • Accomplished global rate limit compliance and zero upstream 429 errors across 300+ requests, as measured by sustaining burst loads up to 61.5 requests/min peak offered rate without queue collapse, by designing a centralized choke-point Gateway featuring dual token buckets, concurrency semaphores, and Earliest Deadline First (EDF) priority scheduling.
  • Accomplished zero-hallucination question answering across 3 arXiv machine learning papers, as measured by 100% citation entailment coverage or truthful refusal, by constructing an end-to-end grounded RAG pipeline with post-generation attribution validation ([{paper_id}v{version}, p.{page}, §{section}]).
  • Accomplished sub-millisecond deadline shedding and cost containment during provider exhaustion, as measured by reducing marginal token expenditure to 0 during quota limits, by engineering an L0–L4 adaptive degradation ladder that automatically sheds expired requests and smoothly falls back to local zero-token extractive search.
  • Links: GitHub Repository
Search Ranking Decay Prediction & Data Contract Pipeline
Python, Scikit-Learn, Random Forest, GroupKFold, SEO Analytics, Data Contracts
Aug 2026 - Oct 2026 More Info
  • Internship Project: Built during my Machine Learning Internship at FlyRank as a core search ranking analytics and client audit platform.
  • Accomplished search ranking position prediction on 30,000+ historical search-log rows, as measured by achieving 84% Precision@50 and 91% Top-20 Precision (a 3.5x lift over the 24% rule-based baseline), by training Random Forest and transformer ranking models.
  • Accomplished audit repeatability and reduced setup overhead across 5 client benchmark runs, as measured by saving 2.0 hours per benchmark run, by building a 4-step automated data contract validation pipeline.
  • Accomplished leak-free model generalizability on out-of-sample client domains, as measured by maintaining 0.00% target feature leakage across 22,006 active content pages, by enforcing strict data contract boundaries and GroupKFold cross-validation.
  • Accomplished model transparency and domain-drift diagnosis, as measured by authoring a comprehensive 38-page engineering evaluation report analyzing model drift, inference latency bottlenecks, and Precision-Recall trade-offs across search verticals.
Kira AI: Real-Time Multimodal Autonomous Desktop Agent
Google Gemini Live API, WebSockets, OpenCV Vision, PyQt6 Cyberpunk HUD, FastAPI
2026 More Info
  • Accomplished low-latency bidirectional multimodal desktop interaction, as measured by sub-second perceived response latency during full-duplex voice streams, by integrating Google Gemini Live API over WebSockets with OpenCV visual screen capture.
  • Accomplished real-time hardware telemetry and UI responsiveness, as measured by maintaining steady 60fps rendering of CPU/GPU/RAM monitoring metrics and live audio visualizer waveforms, by architecting a multi-threaded PyQt6 canvas and async worker threads.
  • Links: GitHub Repository
Jan Setu AI: Civic Technology & Automated Dispatch Platform
Kotlin, Jetpack Compose Android, Gemini Multimodal Vision, Python, FastAPI
2026 More Info
  • Accomplished automated civic hazard triaging and municipal routing, as measured by instant automated severity scoring and zero manual sorting, by engineering an Android application with Kotlin and Jetpack Compose powered by Gemini Multimodal Vision and FastAPI.
  • Accomplished spatial report deduplication, as measured by clustering overlapping citizen hazard reports into single prioritized repair tickets, by implementing GPS spatial clustering and metadata verification on the backend.
  • Links: App Studio Demo • GitHub
FloodCast AI: Satellite ML Flood Hazard Prediction & Early Warning
Python, Scikit-Learn, FastAPI, React 19, Leaflet, Tailwind CSS, Satellite Earth Observation
2026 More Info
  • Accomplished India-wide satellite machine learning flood hazard prediction and inundation early warning across 5 major river basins, as measured by generating high-confidence flood hazard alerts 4 to 5 days prior to peak inundation.
  • Accomplished high-precision spatio-temporal floodplain risk scoring, as measured by training ensemble ML models on multi-source satellite observation data and historical precipitation records to empower disaster management authorities with preemptive lead time.
  • Accomplished interactive geospatial visualization, as measured by sub-second interactive tile rendering across regional flood plains, by engineering a Leaflet map UI connected to a lightweight FastAPI microservice.
  • Links: GitHub Repository
CP-Vault: Competitive Programming Archivist Extension
TypeScript, React, Chrome Extension API (MV3), GitHub API, Express
Aug 2026 More Info
  • Accomplished automated multi-platform solution synchronization to GitHub, as measured by 0 duplicate commits and zero lost submissions across 4 major platforms (LeetCode, Codeforces, CodeChef, HackerRank), by developing a Manifest V3 Chrome extension with SHA-256 deduplication and offline queuing.
  • Accomplished automated repository documentation, as measured by generating structured markdown READMEs with time/space complexity tags and difficulty badges for every accepted submission, by integrating the GitHub REST API with an atomic single-commit workflow.
  • Links: GitHub Repository
Campus Buddy: Student Peer-to-Peer Marketplace
React, Capacitor, Tailwind CSS, Node.js, Express, MongoDB, JWT
Jun 2026 More Info
  • Accomplished a secure, mobile-first peer-to-peer campus marketplace, as measured by sub-50ms database query latency under tested workloads and authenticated student access, by building a full-stack cross-platform app with React, Capacitor, Node.js, Express, and MongoDB Atlas.
  • Accomplished authenticated community trust and reduced asset load times, as measured by 100% verified student email gating and image payload compression, by implementing domain-specific JWT auth and Cloudinary media optimization.
  • Links: GitHub Repository
Gradastra: Higher Education Intelligence Platform
Next.js, TypeScript, Tailwind CSS, Peer Guidance Data Model
Nov 2026 More Info
  • Accomplished an unbiased higher education guidance system, as measured by eliminating commercial agent commission bias (0% commission), by developing an open student-to-senior mentorship platform with side-by-side comparison matrices and AI-assisted scholarship discovery.
  • Accomplished data-driven college shortlisting, as measured by matching applicants with verified senior mentors across leading global institutions based on entrance exam percentiles and career interests.
  • Links: GitHub Repository
YTVidMate: Asynchronous Video & Audio Processing Engine
Next.js, TypeScript, Node.js, Express, FFmpeg, Redis, BullMQ
2026 More Info
  • Accomplished high-performance asynchronous media processing and extraction across 7 platforms, as measured by supporting 6 video resolutions (360p–4K) and high-fidelity MP3/M4A extraction up to 320 kbps.
  • Accomplished resilient background job concurrency without memory exhaustion, as measured by designing an asynchronous processing pipeline with 3 concurrent worker slots, a 30-job Redis queue, and per-IP token bucket rate limiting.
  • Links: GitHub Repository
Amazon ML Challenge 2026: Multimodal Feature Extraction
PyTorch, Multimodal Deep Learning, Transformers, Computer Vision, Scikit-Learn
Sep 2026 More Info
  • Accomplished large-scale multimodal feature extraction in the national Amazon ML Challenge 2026, as measured by zero data pipeline bottlenecks across multi-gigabyte product image datasets, by building deep learning representations in PyTorch.
  • Accomplished efficient distributed batch processing and model inference, as measured by achieving high throughput on dense feature extraction pipelines with zero GPU starvation, by designing custom image pre-processing and parallel data loaders.
  • Links: GitHub Repository
Expense Tracker: Personal Finance Manager
HTML5, CSS3, JavaScript (ES6+), Firebase Realtime Database, Netlify
2025 - 2026 More Info
  • Accomplished lightweight personal financial tracking for students, as measured by real-time visual expenditure breakdown across semester categories with zero data loss, by developing a single-page web app with Firebase Realtime Database and Netlify.
  • Links: Live Demo
Real-Time AI Object Detection & Classification
TensorFlow.js, OpenCV, HTML5 Canvas, Vanilla JavaScript
2025 More Info
  • Accomplished real-time client-side object detection, as measured by achieving low-latency edge inference directly in the browser with zero server dependency, by optimizing TensorFlow.js and HTML5 Canvas processing on live webcam feeds.
  • Links: Live Demo
AI Language Translator & Context-Aware Editor
Python, HuggingFace Spaces, Streamlit, Transformers
2025 More Info
  • Accomplished nuance-preserving multilingual translation, as measured by context-aware grammatical refinement across multiple language pairs, by deploying transformer language models on HuggingFace Spaces with Streamlit.
  • Links: HuggingFace Space

Skills & Platforms

Languages & Core Systems
Python, C++, C, TypeScript, JavaScript, SQL
Machine Learning & AI
PyTorch, Scikit-Learn, Transformers, Hugging Face, OpenCV, Grounded RAG, LLM Gateways, Vector Embeddings
Backend & Architecture
Node.js, Express.js, FastAPI, RESTful APIs, WebSockets, Distributed Caching (Redis), Asynchronous Task Queues, Microservices
Cloud, DevOps & Infra
Amazon Web Services (AWS), Microsoft Azure, Google Cloud (GCP), Docker, Kubernetes, Jenkins, CI/CD Pipelines, Linux / UNIX
Databases & Storage
PostgreSQL, MongoDB, Redis, Vector Databases, Data Contracts, Query Optimization
Quantitative Modeling
Stochastic Modeling, Algorithmic Backtesting, Financial Time-Series, Risk Budgeting (Sharpe, Sortino, Drawdown)

Education

Bachelor of Technology in Computer Science & Engineering
2025 - Present
Computer Science & Engineering
India
More Info
  • Computer Science: Data Structures & Algorithms, Object-Oriented Programming, Operating Systems, Database Management Systems, Computer Networks, Software Engineering.
  • Mathematics & AI: Machine Learning, Deep Learning, Linear Algebra, Probability & Statistics, Discrete Mathematics.
  • Activities: Open-source software development, competitive programming, and technical mentorship.
Diploma in Computer Science & Engineering
2022 - 2025
Computer Science & Engineering
India
More Info
  • Core Foundations: Object-Oriented Programming (C++/Java), Data Structures & Algorithms, Computer Architecture.
  • Applied Systems: Database Management Systems, Operating Systems, Linux Shell Scripting, Software Engineering.
Senior Secondary Education
2012 - 2022
CBSE Class I - X
India
More Info
  • Core Subjects: Physics, Chemistry, Mathematics, and Computer Science.
  • Active participant in inter-school science exhibitions and programming hackathons.