Brijesh Kumar

I'm a

ML Engineer

ML engineer and full-stack developer with 3+ years shipping production AI — RAG pipelines, LLM fine-tuning, computer vision, and cloud infrastructure at scale. Published in Springer Nature. 7+ hackathon wins. Systems live in production serving 10K+ users. M.S. CS @ ASU, 4.0 GPA.

Projects

25+

Total

Awards

Followers

4K+

Tech Stack

55+

Domains

<AI/>

{ data }

[]

()

Skills

Explore my technical skills organized by category

Skills

Languages & Core

Proficient in Python, C++, Go, Java, TypeScript, JavaScript, SQL, Bash, and R — the foundation for building everything from ML pipelines to distributed systems.

Python
C++
Go
Java
TypeScript
JavaScript
SQLSQL
Bash
R

Experience

Software Developer
Current Role
Part Time
EdPlus at Arizona State University
📍 Mission District, San Francisco, CA 94103
Sep 2025 – Present
Building agentic RAG systems and AWS cloud infrastructure to power adaptive learning platforms serving 10K+ enterprise users.
Tech Stack
Python
CLIP
FAISS
Whisper
LangChain
AWS Bedrock
SageMaker
Terraform
GitHub Actions
Key Achievements
Developed agentic RAG system with ReAct-style LLM reasoning, ingesting course documents into FAISS and cutting content prep time 55%.
Engineered AWS Bedrock and SageMaker LLM pipelines for adaptive tutoring and quiz generation, improving engagement 28% for 10K users.
Built CLIP discovery pipeline indexing multi-modal embeddings into FAISS, enabling sub-150ms semantic retrieval across 500K+ segments.
Provisioned Terraform IaC on AWS, eliminating environment drift and cutting infrastructure provisioning time 60% across all team envs.
Software Developer
Current Role
Part Time
EdPlus at Arizona State University
📍 Mission District, San Francisco, CA 94103
Sep 2025 – Present
Building agentic RAG systems and AWS cloud infrastructure to power adaptive learning platforms serving 10K+ enterprise users.
Tech Stack
Python
CLIP
FAISS
Whisper
LangChain
AWS Bedrock
SageMaker
Terraform
GitHub Actions
Key Achievements
Developed agentic RAG system with ReAct-style LLM reasoning, ingesting course documents into FAISS and cutting content prep time 55%.
Engineered AWS Bedrock and SageMaker LLM pipelines for adaptive tutoring and quiz generation, improving engagement 28% for 10K users.
Built CLIP discovery pipeline indexing multi-modal embeddings into FAISS, enabling sub-150ms semantic retrieval across 500K+ segments.
Provisioned Terraform IaC on AWS, eliminating environment drift and cutting infrastructure provisioning time 60% across all team envs.
ML Research Aide
Current Role
Part Time
ASU Biodesign Institute
📍 Mission District, San Francisco, CA 94103
Jan 2025 – Present
Leading ML research on biomedical data at ASU's Biodesign Institute, applying RAG pipelines and deep learning to transform clinical research outcomes.
Tech Stack
Python
FAISS
LangChain
U-Net
ARIMA
Random Forest
Snowflake
AWS SageMaker
PyTorch
SQL
Key Achievements
Integrated RAG-LLM pipeline over clinical embeddings via FAISS, lifting biomarker discovery precision 21% on held-out study cohorts.
Constructed Python and Snowflake ETL on 1.2 TB microbiome data, automating feature engineering and cutting ML preparation time 65%.
Forecasted clinical outcomes via ARIMA and Random Forest, achieving 92% accuracy and cutting prediction error 31% across all domains.
Applied U-Net segmentation on 10 TB histopathology slides, extracting 17 high-signal biomarker features across 5 research teams.
ML Research Aide
Current Role
Part Time
ASU Biodesign Institute
📍 Mission District, San Francisco, CA 94103
Jan 2025 – Present
Leading ML research on biomedical data at ASU's Biodesign Institute, applying RAG pipelines and deep learning to transform clinical research outcomes.
Tech Stack
Python
FAISS
LangChain
U-Net
ARIMA
Random Forest
Snowflake
AWS SageMaker
PyTorch
SQL
Key Achievements
Integrated RAG-LLM pipeline over clinical embeddings via FAISS, lifting biomarker discovery precision 21% on held-out study cohorts.
Constructed Python and Snowflake ETL on 1.2 TB microbiome data, automating feature engineering and cutting ML preparation time 65%.
Forecasted clinical outcomes via ARIMA and Random Forest, achieving 92% accuracy and cutting prediction error 31% across all domains.
Applied U-Net segmentation on 10 TB histopathology slides, extracting 17 high-signal biomarker features across 5 research teams.
Software Engineer
AWL Metaverse Pvt. Ltd.
📍 India
Mar 2024 – Dec 2024
Engineered scalable ML and NLP pipelines on AWS, building intelligent systems that automate workflows and deliver real-time analytics.
Tech Stack
Python
AWS SageMaker
LLaMA2
Transformers
spaCy
RAG
JWT
OAuth2
Pydantic
Power BI
Key Achievements
Generated AI summaries via Transformers, spaCy, and LLaMA2 RAG pipelines, cutting manual reporting effort 50% across all operations.
Secured all API endpoints with JWT and OAuth2 plus Pydantic schema checks, reducing invalid request errors 80% across service layers.
Built end-to-end Python ML pipelines on AWS SageMaker with monitoring, improving model AUC by 0.08 and cutting false alerts 20%.
Software Engineer
AWL Metaverse Pvt. Ltd.
📍 India
Mar 2024 – Dec 2024
Engineered scalable ML and NLP pipelines on AWS, building intelligent systems that automate workflows and deliver real-time analytics.
Tech Stack
Python
AWS SageMaker
LLaMA2
Transformers
spaCy
RAG
JWT
OAuth2
Pydantic
Power BI
Key Achievements
Generated AI summaries via Transformers, spaCy, and LLaMA2 RAG pipelines, cutting manual reporting effort 50% across all operations.
Secured all API endpoints with JWT and OAuth2 plus Pydantic schema checks, reducing invalid request errors 80% across service layers.
Built end-to-end Python ML pipelines on AWS SageMaker with monitoring, improving model AUC by 0.08 and cutting false alerts 20%.
Software Developer
Quicket Solutions
📍 India
Mar 2023 – Mar 2024
Architected high-performance APIs and cloud infrastructure, earning the Best Intern Award for cutting API response time 35% and achieving 99.9% uptime.
Tech Stack
React
Spring Boot
PostgreSQL
pgvector
AWS EC2
RDS
CloudFront
Redis
REST APIs
Key Achievements
Configured pgvector semantic search on 100K+ PostgreSQL records, achieving sub-200ms similarity queries and 99.9% SLA uptime at scale.
Shipped 10+ REST APIs with N+1 query fixes via indexed joins, cutting average API response time 35% — earned the Best Intern Award.
Deployed AWS infrastructure with EC2, RDS multi-AZ, and CloudFront; caching and failover improved throughput 40% at 99.9% uptime.
Software Developer
Quicket Solutions
📍 India
Mar 2023 – Mar 2024
Architected high-performance APIs and cloud infrastructure, earning the Best Intern Award for cutting API response time 35% and achieving 99.9% uptime.
Tech Stack
React
Spring Boot
PostgreSQL
pgvector
AWS EC2
RDS
CloudFront
Redis
REST APIs
Key Achievements
Configured pgvector semantic search on 100K+ PostgreSQL records, achieving sub-200ms similarity queries and 99.9% SLA uptime at scale.
Shipped 10+ REST APIs with N+1 query fixes via indexed joins, cutting average API response time 35% — earned the Best Intern Award.
Deployed AWS infrastructure with EC2, RDS multi-AZ, and CloudFront; caching and failover improved throughput 40% at 99.9% uptime.

Featured Projects

Click the arrow or press space to explore each project

Ghosty — Voice-First Conference Companion (Kiro Spark Winner 🏆)

Won Kiro Spark Challenge 2026 (AWS-sponsored, 24hr) — voice memo → Claude-extracted contact card with transparent Connection Value Score in 15 seconds.
Score engine factors role seniority, company tier, career relevance, and recency; decays 10%/week to surface urgent follow-ups automatically.
Expo + expo-av + Whisper transcription pipeline generates personalized LinkedIn/email drafts referencing the actual conversation — not templates.

Tech Stack

React Native

Expo

TypeScript

OpenAI Whisper

Claude (Anthropic)

Supabase

expo-av

Tags

Mobile AI

Voice AI

Hackathon Winner

GenAI

View on GitHub

Aria — Voice-First Multi-Agent AI Desktop System

5 LangGraph agents (Commerce, Coding, Desktop, General, Docs) with sub-200ms routing lifted task resolution 40% over single-agent baseline.
Full-duplex voice pipeline: Deepgram STT → LangGraph supervisor → Cartesia TTS streaming, delivering <1s end-to-end response latency at scale.
Elasticsearch BM25+RRF hybrid search cut query latency 60% and raised retrieval precision 35%; auto-scales on AWS ECS/Fargate.

Tech Stack

LangGraph

Node.js

TypeScript

Deepgram

Cartesia

Elasticsearch

Playwright

Socket.io

AWS ECS

Tags

Multi-Agent AI

Voice AI

Browser Automation

Real-Time Systems

View on GitHub

RetailForecastAI — Production MLOps Demand Forecasting

End-to-end MLOps: Apache Beam ETL → BigQuery ML ARIMA_PLUS → async FastAPI serving 1,115 parallel models with 24.81M predicted units.
RAG narrative layer (LangChain + FAISS + LLaMA 3.3-70B) auto-generates plain-English per-store summaries via GET /v1/narrative/{store_id}.
Keyless CI/CD via GitHub Actions + GCP Workload Identity Federation: lint → type-check → pytest → Docker → Cloud Run with smoke-test gate.

Tech Stack

Python

Apache Beam

PySpark

BigQuery ML

FastAPI

asyncpg

LangChain

FAISS

Streamlit

Docker

GCP

Tags

MLOps

Time Series

Production AI

GCP

CI/CD

View on GitHub

SlackMind — Production AI Agent with Memory & 59 Tools

Semantic search across 10K+ indexed Slack messages via ChromaDB + OpenAI embeddings (1536-dim) hits 95.3% relevance accuracy in <2s response.
59 tools across GitHub (26), Notion (21), Slack (12) via MCP Protocol — creates issues, queries databases, schedules messages from one chat interface.
mem0.ai long-term memory with GPT-4 Mini fact extraction achieves 98.7% retention across 100+ simultaneous user conversations.

Tech Stack

TypeScript

Node.js

Claude

OpenAI

ChromaDB

mem0.ai

Slack Bolt.js

MCP Protocol

Docker

Tags

Multi-Agent AI

RAG

Enterprise Automation

Real-Time Systems

View on GitHub

StyleNova AI — Vision-Language Adaptive Fashion Recommender

Hybrid scorer fusing CLIP ViT-B/32 (512-dim), TF-IDF content signals, and user-based kNN collaborative filtering achieves Precision@10 of 0.73.
Online preference learning via EMA (λ=0.3) with ε-greedy exploration improves recommendation relevance 32% after just 5 user interactions.
Pre-computed CLIP embeddings cached as SQLite BLOBs deliver sub-200ms p95 API response for 10K+ concurrent users — zero cold-start latency.

Tech Stack

Python

PyTorch

CLIP

FastAPI

scikit-learn

Next.js

TypeScript

Prisma

Zustand

Framer Motion

Tags

Computer Vision

Recommendation Systems

Deep Learning

E-commerce

View on GitHub

QuarkCache — Redis-like In-Memory Store Built in C++17

O(1) LRU eviction via doubly-linked list + unordered_map; std::list::splice() promotes nodes with zero allocation — ~80K ops/sec on local TCP.
Non-blocking TCP server using epoll/kqueue in edge-triggered mode handles 10K+ concurrent connections on a single thread, no context-switch overhead.
Hybrid TTL expiry: lazy O(1) deletion on access + background sweep every 5s via steady_clock, mirroring Redis's production expiry strategy.

Tech Stack

C++17

epoll/kqueue

FastAPI

Python

Next.js

TypeScript

Docker

CMake

Tags

Systems Programming

Low-Latency

Distributed Systems

DevOps

View on GitHub

KubeGuard — AIOps Kubernetes Anomaly Detection & Self-Healing

Hybrid LSTM + LightGBM on Prometheus 1Hz pod metrics predicts anomalies before breach, cutting CrashLoopBackOff events 38% vs threshold alerting.
Go Prometheus exporter surfaces kubeguard_predicted_memory_ratio as a custom K8s metric, enabling HPA auto-scaling via Prometheus Adapter.
Gemini 1.5 Pro generates targeted kubectl remediation commands on confirmed anomalies via LangChain, cutting MTTR 40%.

Tech Stack

Python

TensorFlow

LightGBM

Kubernetes

Prometheus

LangChain

Gemini 1.5 Pro

Docker

Tags

AIOps

Kubernetes

Anomaly Detection

DevOps

LLM Remediation

View on GitHub

EmotiSphere — AI 3D Avatar Conversational Platform

Three.js avatar with 26 MorphTarget blend shapes and PBR shading sustains 60 FPS — won Top 3 at ASU Hackathon for phoneme-accurate lip sync.
Audio → FFmpeg → Rhubarb v1.14 → 22-viseme JSON → MorphTargets pipeline achieves <5ms audio-to-viseme precision at runtime.
Llama 3.1 8B via Ollama delivers 87% intent accuracy and 82% emotion detection; Node.js + MongoDB + Clerk JWT powers persistent sessions.

Tech Stack

React

Three.js

TypeScript

Node.js

Ollama (Llama 3.1)

ElevenLabs

MongoDB

Clerk JWT

FFmpeg

Tags

GenAI

3D Avatar

Real-Time AI

Hackathon Winner

View on GitHub

EduMorph — Self-Improving AI Educational Platform

Azure Event Hubs → Stream Analytics → Synapse Analytics pipeline ingests 1,000+ events/sec, delivering 92% content accuracy for 10K+ learners.
Power BI dashboards via DirectQuery to Synapse surface concept difficulty, format effectiveness, and cohort engagement trends in real time.
GPT-4o adaptive loop analyses telemetry and automatically regenerates improved content — boosting engagement 85% and retention 78%.

Tech Stack

Next.js 14

TypeScript

Azure Event Hubs

Azure OpenAI

Groq

Synapse Analytics

Power BI

Recharts

Zustand

Tags

GenAI

EdTech

Big Data

Azure

Scalable Systems

View on GitHub

StudySlice AI — Lecture-to-Study-Clip Intelligence System

Gemini 2.5 Flash + AWS Transcribe processes 3+ hr Harvard CS50 lectures via 2-min overlapping windows with 85%+ concept identification accuracy.
Diversity-enforced clip selection across 5 concept types (Definition, Example, Process, Summary, Question) yields 10 focused 40-second study clips.
Production AWS stack (S3 + Transcribe + EC2 + IAM roles) cuts lecture review time ~95% — from 3 hours down to under 7 minutes.

Tech Stack

Python

Flask

Google Gemini

AWS S3

AWS Transcribe

FFmpeg

Next.js

TypeScript

Framer Motion

Tags

EdTech

Video Intelligence

AWS

NLP

View on GitHub

Swiggy Market Intelligence Engine — Strategic Food Analytics

4-signal City Expansion Index (Revenue 30%, Rating 25%, Density 25%, Diversity 20%) scored every city 0–100 across 197,430 orders.
BCG-style Menu Intelligence Matrix tiered 1,200+ restaurants into Champion/Healthy/At Risk/Critical — flagging ~15% for platform intervention.
ARIMA forecasting confirmed Pareto effect (~20% cities = ~80% revenue); delivered 7-tab Streamlit dashboard + 13-sheet Excel KPI report.

Tech Stack

Python

Pandas

Plotly

Streamlit

SQLite

SciPy

Statsmodels

Scikit-learn

openpyxl

Tags

Data Analytics

Business Intelligence

EDA

Market Strategy

View on GitHub

TrustMed-AI — Production Multi-Agent Medical RAG System

LLaMA2 with context-only guardrails and multi-agent routing on AWS ECS achieves 100% query accuracy on held-out medical test set at 99% uptime.
Hybrid RAG combining Sentence-BERT reranking and ChromaDB vector search lifts medical retrieval precision 75% over keyword-only baseline.
ThreadPoolExecutor parallel agents with utility-based scoring cuts query latency 12s → 4s — 3× speedup without any accuracy degradation.

Tech Stack

Python

LangChain

LLaMA2

ChromaDB

FastAPI

AWS ECS

Docker

Sentence-BERT

ReactJS

Tags

GenAI

Multi-Agent AI

RAG

Healthcare AI

Production ML

View on GitHub

Smart Contact Manager — AI-Powered Network Intelligence CRM

Proprietary ML scoring (0–100) with recency decay and channel weighting (Meeting 10×, Call 7×, Email 4×) recalculates in real time on every interaction.
Claude 3.5 Sonnet (200K context) handles network queries, email drafts, and note summarisation at <50ms Convex DB latency.
Levenshtein-based fuzzy deduplication + dormant-contact detection via score decay covers 100K+ records with zero manual review.

Tech Stack

Next.js 16

TypeScript

Convex

Claude 3.5 Sonnet

Clerk

Recharts

Framer Motion

Tailwind CSS

Tags

GenAI

Full-Stack

CRM

Real-Time Systems

View on GitHub

JobHunt — AI-Powered Career Platform

RAG-based resume-job matching with vector embeddings boosts application success rates 40% for active job seekers.
Resilient microservices handle 180 RPS with zero errors under sustained load — auth, search, and job-matching all independently scalable.
Kafka-driven Java scraping pipeline expands the job database 3× continuously with zero manual data entry.

Tech Stack

Next.js

Spring Boot

MongoDB

Java

RAG

Vector Embeddings

Microservices

Kafka

Tags

GenAI

Full-Stack

Automation

Microservices

View on GitHub

Spotify Top Songs Analytics Dashboard

Automated Spotify API ETL pipeline cuts manual data prep 80%, enabling daily refresh across 50K+ tracks with zero intervention.
Statistical EDA uncovered 5 viral trend patterns; Scikit-learn model predicts playlist inclusion at 0.92 R² score.
Tableau and Power BI dashboards surface engagement trends consumed by 3 content strategy teams for data-driven curation decisions.

Tech Stack

Power BI

Python

SQL

DAX

Spotify API

Scikit-learn

Tableau

Tags

Data Analytics

Visualization

Machine Learning

ETL

View on GitHub

Education

Arizona State University

Currently Pursuing

Hover for details

Tempe, AZ

Jan 2025 – Present

4.00 / 4.00

CGPA

18 / 18

Credits Earned

Projects

Hackathons Won

M.S. in Computer Science

Key Coursework:

Statistical Machine Learning
Semantic Web Mining
Data Mining
Data Processing at Scale

Vellore Institute of Technology

Completed

Hover for details

Chennai, India

Jul 2019 – May 2023

4.00 / 4.00

CGPA

160 / 160

Credits Earned

Projects

Hackathons Won

B.Tech in Computer Science (AI/ML Specialization)

Key Coursework:

Machine Learning
Deep Learning
NLP
Data Structures and Algorithms
Computer Vision

Awards & Recognition

Winner — Kiro Spark Challenge 2026

Kiro / AWS · Arizona State University

Mar 2026

Won a 24-hour AWS-sponsored hackathon with Ghosty — a voice-first mobile app that turns conference conversations into a transparent career pipeline. 15-second voice memo → AI-extracted contact card with Connection Value Score → personalized follow-up draft.

Runner-up — SunHacks 2025 Education Track

SunHacks · Arizona State University

Nov 2025

Placed 2nd in the Education Track with StudySlice AI — an AI pipeline (Gemini + AWS Transcribe) that converts 3-hour lectures into 10 focused 40-second study clips, cutting review time 95%.

Top 3 — EmotiSphere Hackathon

Hackathon

Nov 2025

Won Top 3 for building a real-time AI 3D avatar with phoneme-precise lip-sync, Llama 3.1 emotional intelligence, and sub-5ms audio-to-viseme synchronization.

Hacktoberfest Super Contributor

DigitalOcean & GitHub

Oct 2025

Recognized as a Super Contributor for exceptional open-source contributions during Hacktoberfest 2025 across multiple repositories.

Finalist — Code for Good 2024

JPMorgan Chase & Co.

Oct 2024

Selected as a finalist among thousands of participants for developing innovative tech solutions addressing real-world social challenges. Earned direct interview opportunity with JPMorgan Chase engineering teams.

Winner — CrozeAI Hackathon

CrozeAI

Jul 2023

Won by architecting an intelligent routing algorithm for NYC bike navigation using advanced pathfinding optimized for safety, efficiency, and real-time traffic patterns.

Top 0.2% — CodeKaze Programming Contest

Coding Ninjas

Feb 2024

All India Rank 315 out of 150,000+ participants, College Rank 3 at VIT. Solved 1,000+ LeetCode problems across advanced algorithms and data structures.