System Online · Polaris / Medhavi
Vinit Jangir
AI/ML & Full-Stack Engineer.
ARCHITECTING SCALABLE INTELLIGENCE.
BRIDGING DATA & KINETIC LOGIC.
Architecting full-stack ecosystems and deploying high-performance AI models. Bridging the gap between raw data and kinetic logic.
AI & ML Systems
RAG architectures, face detection, predictive modeling — from research to production.
Cloud Infrastructure
Docker, AWS, CI/CD — systems engineering at undergraduate level.
Full-Stack Dev
React, FastAPI, Go — clean architecture that performs under real-world load.
Open Source
Active upstream contributor across p5.js and open-source ecosystems — 100K+ line codebases.
Technologies I work with daily
Engineering
Intelligence
& Scalable
Systems.
HOST: Vinit Jangir
CORE: AI/ML · Distributed Systems · Cloud
BASE: Polaris School of Tech (Medhavi Skills Univ)
FOCUS: 1,500 Concurrency · High-Throughput Engines
I'm Vinit Jangir — an AI/ML specialist and full-stack engineer at Polaris School of Technology (Degree by Medhavi Skills University). I specialize in building intelligent systems at the intersection of machine learning and production-grade infrastructure. My work spans the entire pipeline: from training predictive models and orchestrating RAG architectures, to containerizing services with Docker and deploying them on cloud infrastructure.
With a keen eye for system design and a deep understanding of algorithms, I architect solutions that don't just work — they scale. Whether it's building a real-time AI proctoring engine or contributing inference improvements to open-source ML libraries, I blend strategy, performance, and clean code to bring ideas to production.
Experience & Milestones.
Track record of technical mentorship, upstream open-source patches, and competitive hackathon leadership.
Technical Mentor & Product Strategist
CAMPUS CONNECT · REMOTE
- • Providing high-fidelity technical roadmaps and pre-college systems architecture guidance to 100+ engineering aspirants.
- • Optimized student advisory workflows by 20% through structured data frameworks and industry gap analysis.
Open Source Core Contributor
VARIOUS ECOSYSTEMS (ESOC) · REMOTE
- • Contributed upstream pull requests to 100K+ line open-source Python ML codebases: expert-knowledge API string representations, in-place return-contract fixes in role mixins, and doctest repairs.
- • Modernized p5.js at the Processing Foundation through deprecated WebGL cleanup. Active contributor across sktime and AI-on-Demand.
Smart India Hackathon (SIH 2026)
MINISTRY OF HOME AFFAIRS (I4C) · PROJECT DRISHTI
Cleared internal selection round and advancing to the National Grand Finale on Problem Statement SIH26184: developing cybercrime predictive dispatch pipelines using LightGBM LambdaMART rankers, Cox survival modeling, and Uber H3 spatial indexing.
B.Tech in Computer Science & AI
POLARIS SCHOOL OF TECHNOLOGY · MEDHAVI SKILLS UNIVERSITY
Specialized curriculum in distributed systems, machine learning engineering, and algorithmic complexity. Solved 250+ algorithm challenges across competitive platforms while architecting production-grade full-stack systems.
Why Python for DSA?
Algorithm Design · Competitive ProgrammingPython isn't my crutch — it's my scalpel. While others debate languages, I leverage Python's expressive syntax to prototype O(n log n) solutions in minutes, then validate them against thousands of edge cases. The result? Cleaner AI logic. When your preprocessing pipeline, your model inference, and your algorithmic optimizations all speak the same language, you eliminate translation overhead and ship faster.
Architectural Mastery.
Precision systems engineering across distributed backends, algorithmic optimization, upstream open-source codebases, and kinetic client interfaces.
DISTRIBUTED BACKEND & INFRASTRUCTURE
Engineered production microservices, asynchronous BullMQ queues, Redis caching, and containerized Docker pipelines. Validated under k6 load tests with zero degradation.
ALGORITHMS & APPLIED ML
Competitive problem-solving across LeetCode and CodeChef. Architecting two-stage LLM routers (8B → 70B), RAG FAISS vector stores, and LightGBM ranking engines.
OPEN SOURCE ECOSYSTEMS
Active upstream contributor across open-source codebases (Python ML & computer vision) and p5.js at Processing Foundation. Navigating strict CI/CD pipelines, doctests, and distributed collaboration.
CLIENT ARCHITECTURE & KINETIC UI
Building responsive web applications, hardware-accelerated 60fps card stacking with Lenis, financial charts with Chart.js, and type-safe component state machines.
What I've built.
Explore my recent systems — engineered to solve real concurrency, reliability, and algorithmic bottlenecks with production-grade code.
Vision
Situation
The exam platform saturated at 500 concurrent students — autosave p95 hit 8 seconds and sessions took 9 seconds to start.
Action
Profiled with k6 and root-caused it: a middleware chain running twice per autosave, a MongoDB write on every request (~130/sec), and a hard-coded connection pool of 10. Rebuilt grading as an async BullMQ pipeline — 202 Accepted, Judge0 workers, Socket.IO result delivery.
Result
1,500 concurrent students on a single instance at 0.00% failures. Autosave p95 8,012 ms → 48 ms (167×), session start 9,192 ms → 27 ms (340×). The breaking point was never located.
Sahayak
Situation
A stranded driver needs the right mechanic, not just the closest one — nearest-first ignores skill, equipment, and current workload.
Action
Own the backend, database, and infrastructure. Built a dispatch engine with eligibility gates (verification, availability, equipment, job cap, 120-second location freshness) over PostGIS and Redis geospatial, then weighted ranking across distance, workload, skill, and rating.
Result
Every dispatch also logs what nearest-only would have picked — a shadow-mode quality benchmark with zero A/B risk to real users.
Cortex
Local Neural AssistantSituation
General-purpose AI assistants forget context between sessions, and routing every message to a large model is slow and expensive.
Action
Built a two-stage LLM router — an 8B classifier tags intent in ~200 ms, then dispatches to a 70B responder, a web-search path, or a vision model. Added RAG long-term memory over FAISS with local embeddings, and a full voice pipeline: Whisper STT, Piper TTS, speech endpointing, and a hands-free wake-word state machine.
Result
A 76K-line local-first assistant with persistent memory, streaming speech, vision, and an Android companion — running entirely on my own machine.
FluxCurrency
FX DashboardSituation
Currency converters show you a number and nothing else — no trend, no context for whether today's rate is good.
Action
Built a real-time dashboard over the Frankfurter API (ECB-sourced) covering 30+ currencies, with 1M/3M/1Y historical charts, a live market ticker, debounced input, and persisted theme.
Result
A deployed, responsive glassmorphic dashboard that turns a one-off conversion into an informed decision.
Drishti
Cybercrime Interception SIH 2026 · Cleared Internal → Grand FinaleSituation
Mule-account networks move stolen money through hundreds of accounts faster than investigators can trace them — policing stays reactive, always one hop behind.
Action
Building the prediction pipeline for Problem Statement SIH26184 (Indian Cybercrime Coordination Centre, Ministry of Home Affairs): a LightGBM LambdaMART ranker for where, a Cox proportional-hazards survival model for when, with SHAP explanations over H3 geospatial indexing.
Result
Cleared internal selection round and advancing to the National Grand Finale — shifting policing from reactive tracing to physical interception.
Contributing to the ecosystem.
Open source isn't a checkbox on my resume — it's how I sharpen my engineering instincts.
I've contributed upstream pull requests to production open-source ML libraries (100K+ LOC) and p5.js at the Processing Foundation. I also actively contribute across sktime, OpenML, and AI-on-Demand.
Navigating unfamiliar 100K+ line codebases at scale, shipping CI-validated patches, and communicating technical decisions asynchronously with distributed teams across time zones. This real-world exposure allows me to write cleaner, more maintainable code and understand the nuances of large ecosystem architectures.
Certifications & Activities.
Smart India Hackathon 2026
Cleared internal selection round and advancing to the National Grand Finale on Problem Statement SIH26184 for the Indian Cybercrime Coordination Centre, Ministry of Home Affairs.
Organizing Committee Member
Managed technical infrastructure for major workshops from Aug 2025 – Jan 2026, ensuring 100% uptime for live sessions and seamless delivery.
Let's build something
intelligent.
I'm open to production engineering internships, GSoC 2026, and open-source fellowships. If you have an ambitious problem to architect, initialize dispatch below.