Let’s talk ↗
Chittagong, Bangladesh → Remote everywhere /// AI · LLM · Agents · Evaluation
Arnob Rizwan Ahmad Arnob Rizwan Ahmad

Arnob Rizwan Ahmad.

AI/LLM engineer building agentic workflows, RAG and evaluation infrastructure, and the backend they run on. Currently ERP Developer at Hirdaramani Bangladesh, previously Agentic AI Engineer at BLOOMTECH and ML Engineer for a US news-intelligence platform. Software Engineering graduate of Universiti Teknologi Malaysia (July 2026, First Class), shipping for clients in the US, Luxembourg, Spain, Bangladesh and Malaysia.

Download Resume ↓ GitHub ↗ LinkedIn ↗
47
Projects On This Page
5
Countries Shipped To
40
Schools On My App
10
AI Agents, One Platform
01

Experience

Full-time and contract roles for teams in Bangladesh, the US and Luxembourg, start to ship.

/01

ERP Developer

Hirdaramani — Bangladesh 🇧🇩
Jul 2026 — Present · Hybrid
  • Developing and maintaining ERP modules supporting production, inventory, and order-management workflows across factory operations.
  • Writing and tuning SQL against large operational datasets, building reports and data integrations between ERP and internal systems.
  • Building the Lab Test Management System (LTMS): ASP.NET Core and Angular over SQL Server, with multi-level approvals, lab workflow, payments and multilingual reports.
  • Working with operations and IT stakeholders to translate process requirements into shipped features.
ERPSQLData IntegrationReporting
/02

Agentic AI Engineer

BLOOMTECH, Inc. — Remote 🇺🇸
Aug 2026 — Oct 2026 · Part-time
  • Built agentic AI workflows in Python — multi-step LLM planning, tool and API integration, and structured output handling.
  • Designed prompt and evaluation loops to measure and improve agent reliability across task categories.
LLM AgentsPythonEvalsTool Use
/03

Application Developer

Moien Delivery — Luxembourg 🇱🇺
Mar 2026 — Jul 2026 · Remote
  • Built the Angular web front-end and a cross-platform React Native mobile app on a Nuxt.js backend, serving three user roles: customers, restaurants, and drivers.
  • Implemented real-time order tracking, driver dispatch, and delivery workflow automation over REST APIs and WebSocket event streams.
  • Owned features from spec → QA → production on weekly release cycles with international stakeholders.
AngularReact NativeNuxt.jsWebSockets
/04

Machine Learning Engineer

AI News HUD (via Mercor) — USA 🇺🇸
Oct 2025 — Mar 2026 · Remote Contract
  • Built and evaluated production PyTorch models for real-time news classification, summarization, and personalized recommendations.
  • Designed evaluation pipelines and ran iterative experiments on live traffic, raising news-classification accuracy from 70% to 78.9% and cutting inference latency 20%.
  • Took ML components from prototype to deployed feature in the platform's core intelligence layer.
PyTorchEval PipelinesNLPProduction ML
/05

Founding Backend Engineer

Framic — Remote
Aug 2025 — Jul 2026 · Remote · ~10 hrs/week
  • First backend hire on an AI-powered storage app — owned architecture decisions and core API design from scratch.
  • Led AI search, real-time file indexing, and scalable gallery/drive systems.
  • Worked directly with founders to ship fast while keeping the system maintainable as the team scaled.
System DesignAI SearchAPIs0→1
/06

Software Engineer (Freelance)

MOMS Mobile Oil Change — USA 🇺🇸
Apr 2026 — Jul 2026 · Remote
  • Redesigned site header/navigation (active routes, hover dropdowns, CTA) and rebuilt three product pages for a live Next.js business site.
  • Led SEO overhaul: dynamic sitemaps, 301 redirects for legacy URLs, and dynamic routing under /locations/[city].
  • Conducted site-wide security and dependency audits, and resolved production-build breakages.
Next.jsReactTypeScriptSEO
/07

Software Engineering Resident

Headstarter — USA 🇺🇸
Sep 2024 — Mar 2025 · Remote
  • Selected for a competitive 6-month US engineering residency; shipped ML and full-stack projects end to end in Python and TypeScript.
  • Designed and trained neural networks using LLM chaining, hyperparameter tuning, and fine-tuning.
  • Built LLM-powered agents on AWS Bedrock / OpenAI APIs and optimized Docker CI/CD on GitHub Actions, cutting image size and build times.
Neural NetsAWS BedrockCI/CDML + Full-Stack
02

Live Projects

Not repos — shipped products. Sites you can open right now and apps you can install on your phone.

Screenshot of the ShipCheck homepage Live · Free Tool · No Login

ShipCheck

shipcheck-gold.vercel.app

Paste a website, an App Store link or a Google Play package and get a scored health check with the exact fix for every finding. Websites are scored across seven categories, with Lighthouse behind the performance numbers. I calibrated it against youhadonejob.dev, and on 26 real sites every score landed within 10 points.

Node.jsServerlessLighthouseWeb SecurityVercel
Open the tool ↗
Screenshot of the Relay homepage Live · Real-Time · Offline-First

Relay

taghyeer-chat-rust.vercel.app

Real-time direct and group chat on Next.js with an offline outbox that survives a reload. Tabs signed in as the same user stay in sync, and since the free-tier backend sleeps, the landing page wakes it up while you read. 91 unit tests, clean lint and type check.

Next.jsTypeScriptSocket.ioTanStack QueryZustand
Open the app ↗
Screenshot of the StockTrade homepage Live · Local LLM · No Login

StockTrade

stocktrade-topaz.vercel.app

Stock social platform that scores community posts for insight quality and rolls reputation-weighted sentiment into a per-ticker Smart Signal, using a local LLM over Ollama instead of a paid API. When Yahoo Finance blocks live prices, it falls back to a timestamped snapshot and says so in a banner.

Next.js 16PrismaOllamaClerkRSS
Visit site ↗
Screenshot of the FF Gamer BD homepage Live · Client Work · Bangladesh 🇧🇩

FF Gamer BD

ffgamerbd.com

Live Free Fire top-up store in Bangladesh on Laravel and Filament. I did the backend integration work: connected a reseller API for automatic order fulfilment and built the payment gateway integration, where every payment callback is checked with the provider before an order is marked paid.

LaravelPHPPayment GatewayWebhooks
Visit site ↗
Screenshot of the PI STAR homepage Live · Client Work · Bangladesh 🇧🇩

PI STAR

pistarbd.com

Bangladeshi streetwear and tech-accessories storefront on Node.js and Express, with bKash, Nagad, Rocket and cash-on-delivery checkout. I was brought in during a production outage, traced it to the hosting setup and a broken dependency build, and got the store serving again.

Node.jsExpressSQLiteIncident Response
Visit site ↗
Screenshot of the Rubicon Liberty Group homepage Live · Product Build · Crypto Markets

Rubicon Liberty Group

crypto-rlg.vercel.app

Digital asset markets platform, built as a self-directed product and not a live financial service. Live pricing for 50 assets from CoinGecko, refreshed every 60 seconds, with market stats, search, gainer and loser filters and a seven-day sparkline per asset. Demo status and risk warnings are stated in the footer.

Next.jsCoinGecko APILive DataVercel
Visit site ↗
Screenshot of the Al-Ibadah homepage Live · Bangla-First · Bangladesh 🇧🇩

Al-Ibadah

alibadahfoundation.com

Bangla-first Islamic reference site: the full Quran with five Bangla translations and ten reciters, 36,432 hadith, daily duas, and prayer times and qibla for all 64 districts. The /ask assistant runs entirely in the browser and refuses rather than invents. Donations go through PipraPay with bKash, Nagad and Rocket, and a Flutter companion app is in closed testing.

React 18ViteTypeScriptStatic Prerender
Visit site ↗
Screenshot of the MOMS Oil Change homepage Live · Client Work · USA 🇺🇸

MOMS Oil Change

momsoilchange.com

Live production site for a US mobile oil change and fleet maintenance business in Philadelphia. I redesigned the header and navigation, rebuilt three product pages, and led an SEO overhaul — dynamic sitemap, 301 redirects for legacy city URLs, and consolidation onto /locations/[city] dynamic routes.

Next.jsReactSEOProduction
Visit site ↗
Screenshot of the Pizza Lab homepage Live · Client Work · Spain 🇪🇸

Pizza Lab

pizzalabgroup.com

Bilingual English and Spanish site for a pizzeria in Estepona, Spain, with a no-code admin where the owner edits the menu, prices, hours and photos and the site updates instantly. I also keep the printed English and Spanish menus in step with the site, most recently adding a new dessert and a scannable Instagram QR code.

Node.jsExpressCMSVercel
Visit site ↗
Screenshot of the SafeCurve homepage Live · Road Safety · Simulation

SafeCurve

safecurve-sim.vercel.app

Interactive Three.js driving simulation backing a road-safety study — drivers experience a blind-curve scenario, react to escalating LED and buzzer warning levels, and rate the system, generating real user-acceptance data. Also packaged with Capacitor into a fully offline 4 MB Android APK.

Three.jsCapacitorAndroidSimulation
Visit site ↗
Screenshot of the Best Auto Rentals homepage Live · Full Platform · 80 Tests

Best Auto Rentals

bestauto-rentals.vercel.app

Full car rental platform — customer booking flow, functional admin dashboard, a provider-agnostic AI concierge with output validation, and an event-driven automation engine. 80 automated tests across 20 suites, with CSP, HSTS and Permissions-Policy headers in production.

Next.js 16React 19Neon PostgresDrizzle
Visit site ↗
Screenshot of the PhotoniX Research homepage Live · Client Work · Research

PhotoniX Research

photonixresearch.com

Production website for a photonics and laser research group — research areas, publications, team profiles, news and recruitment, with a motion-driven front end.

WebMotion UIProduction
Visit site ↗
Screenshot of the JobPortal homepage Live · Full-Stack · Product

JobPortal

jobportal-blush.vercel.app

Job application platform with company profiles, job listings, a candidate leaderboard and a blog — built to make the application journey measurable rather than a black box.

Full-StackReactVercel
Visit site ↗
Installable App · Android + iOS

ZeroWaste Hub

Signed APK & IPA on GitHub Releases

Food-rescue app for Malaysia connecting donors, recipients and volunteer riders — Gemini vision grades dish quality and halal status, an on-device TFLite classifier caches its own model, with Checkout.com / TNG eWallet payments and four languages.

↓ APK · 79 MB↓ iOS IPA · 25 MB
FlutterGemini VisionTFLite
Download build ↗
Screenshot of the Moien Delivery homepage Live on Stores · Luxembourg 🇱🇺

Moien Delivery

moiendelivery.lu · Google Play & Apple App Store

Food delivery platform for Luxembourg serving customers, restaurants and drivers. I built the Angular web front-end and a cross-platform React Native app, with real-time order tracking, driver dispatch and delivery workflow automation over REST APIs and WebSocket event streams.

AngularReact NativeNuxt.jsWebSockets
Support Welfare Companies app screens on Google Play Live on Google Play · Joint Project

Support Welfare Companies

Google Play · Android · source repo private

A joint project, live on Google Play, that ranks employers on how they treat their workers. A Flutter app for the public and a React + TypeScript reviewer panel share one Firebase backend. Companies are scored on five welfare criteria, and every survey answer is checked against public evidence before it publishes.

FlutterReact TSFirebaseGoogle Sign-InFirestore Rules
03

Applied LLM Systems

Fifteen systems an AI engineer actually gets asked to build — retrieval, evaluation, guardrails, routing, agents. Every one is hosted and interactive: open a card, type your own input, and the real code runs and shows you its actual output. 494 tests, no API key anywhere. applied-llm-systems.vercel.app · source

Screenshot of the applied-llm-systems homepage Live · 15 Interactive Demos · 494 Tests · Zero Paid APIs

applied-llm-systems

applied-llm-systems.vercel.app · github.com/Arnobrizwan/applied-llm-systems

Fifteen production-shaped LLM systems in one codebase on a dependency-free Python core. Every demo and test runs offline with no API key, and a CI gate fails the build if any runtime file imports a third-party package. Point one environment variable at Ollama or any OpenAI-compatible endpoint and the same code runs against a real model.

PythonRAGEvalsAgentsGuardrailsObservability
Open the live demos ↗
System 01 · Live · Run it now

Production RAG Pipeline

Ingestion, chunking, hybrid BM25 + vector search fused with reciprocal rank fusion, reranking, and answers where every citation is validated against a retrieved chunk before it ships. Reranking lifted recall@1 from 0.67 to 0.93 and MRR to 0.967.

Hybrid SearchRRFCitations
↗
System 02 · Live · Run it now

Structured Output Engine

A JSON Schema validator, extractor and repair chain written from scratch, with validation errors fed back into the retry. With every response deliberately corrupted, success went from 50% at one attempt to 88% at three, half of it repaired locally with no second model call.

JSON SchemaRepairRetries
↗
System 03 · Live · Run it now

Context Assembly Service

Token budgeting across memory, documents, tools and history, with per-section minimums, compression strategies and a reserved completion budget. Naive concatenation overran an 800-token window by 426 tokens; the assembler overflowed none and kept every high-priority source.

Token BudgetCompressionReceipts
↗
System 04 · Live · Run it now

LLM Evaluation Harness

Golden dataset, deterministic scorers, a rubric and pairwise LLM judge, bootstrap confidence intervals, and a CI gate that blocks a merge on regression. Pairwise judging showed 78.3% position bias, so those verdicts are recorded as ties rather than trusted.

LLM-as-JudgeBootstrap CICI Gate
↗
System 05 · Live · Run it now

Semantic Cache Layer

An embedding cache with TTL, LRU eviction and namespace isolation. The paraphrase and near-miss sets turned out not to be separable by any threshold, so a salience guard compares numbers, entities and negations: it removed all five false hits without losing a real one.

EmbeddingsCache SafetyLRU
↗
System 06 · Live · Run it now

Model Routing Gateway

Complexity scoring across explainable features, three tiers, per-tenant overrides, budget guards and a circuit-broken fallback chain. 55.7% cheaper than always-large across 44 requests, and it served all 44 through a total outage of the top tier.

RoutingCost ControlCircuit Breaker
↗
System 07 · Live · Run it now

Multi-Tenant LLM API

Salted-hash API keys, token-bucket rate limits, reserve-then-reconcile token budgets and namespaced storage, served over plain http.server. A cross-tenant read returns a byte-identical 404 to a fictional id, so the API is not an existence oracle.

Multi-TenancyQuotasIsolation
↗
System 08 · Live · Run it now

Fine-Tuning Pipeline

Synthetic data generation, dedupe and stratified splits, then LoRA implemented from scratch: frozen base weights, a trained low-rank pair, alpha/r scaling and a merge. 1,048 trained parameters against 1,542 for a full fine-tune, at the same accuracy.

LoRAFrom ScratchBenchmarks
↗
System 09 · Live · Run it now

Agent Memory System

Working, episodic and semantic memory with summarisation on overflow, contradiction handling that supersedes rather than deletes, and decay-based eviction. A fact stated at turn 2 is still recalled at turn 32 inside a 300-token budget where the baseline fails.

MemoryCompressionEviction
↗
System 10 · Live · Run it now

Guardrails Middleware

Weighted injection detection, reversible PII redaction with Luhn-checked card numbers and Bangladeshi and Malaysian phone formats, output filters and a policy engine that records which rule fired. 140 PII decisions with no false positives or negatives, at 0.1ms overhead.

PIIInjectionPolicy
↗
System 11 · Live · Run it now

Streaming Response Infrastructure

Server-sent events with heartbeats, bounded-queue backpressure under real TCP flow control, cancellation, and Last-Event-ID resume. A connection killed mid-stream resumes to 277 contiguous ids with zero duplicates and no regeneration.

SSEBackpressureResume
↗
System 12 · Live · Run it now

Prompt Versioning and A/B

Content-hash versions, environment labels, hash-based sticky traffic splitting, outcome tracking and a two-proportion z-test promotion gate written in pure Python. It refused to promote at n=40 despite a significant p-value, then promoted at n=600.

VersioningA/B TestingRollback
↗
System 13 · Live · Run it now

LLM Observability Stack

Nested tracing with token, latency and cost attribution, privacy-safe capture, EWMA anomaly detection and alert cooldown. Seven alert events became three pages, no span carried raw prompt text, and 36% of calls turned out to be 98% of the bill.

TracingCostAlerting
↗
System 14 · Live · Run it now

Tool-Calling Framework

Schemas derived from type hints, a discoverable registry, argument coercion, per-tool circuit breakers and sandboxed execution with a safe AST evaluator. Twelve sandbox escape attempts rejected, including dunder walks and huge exponents, with real arithmetic still working.

Tool UseSandboxingSchemas
↗
System 15 · Live · Run it now

Self-Correcting RAG Agent

A bounded loop that rewrites the query, critiques its own retrieval, escalates through a fallback ladder and abstains below a confidence floor. 16 of 19 answerable questions right against 13 single-shot, and it refuses 5 of 8 unanswerable ones instead of inventing an answer.

AgentsSelf-CorrectionAbstention
↗
04

Projects

Thirteen builds I'd defend in a code review — agents, LLM evaluation infrastructure, MLOps, and mobile. The rest of the archive lives on GitHub.

AutoML · Agents

Autonomous ML Agent

An LLM planner proposes modeling strategies under a time budget, Optuna tunes XGBoost/LightGBM/sklearn, ensembles get built, and SHAP explains the winner — shipped with model cards, a FastAPI inference service, an automated test suite, and CI/CD pushing to Docker Hub.

OptunaSHAPCI/CD
↗
LLM Evals · Infra

Self-Improving Prompt API

CI/CD for prompts — a semver-versioned registry where each improvement cycle runs failure analysis, generates candidates, A/B-evaluates them with a blinded LLM judge, and auto-promotes winners per YAML rules. Multi-dimensional scoring across OpenAI and Anthropic providers.

FastAPILLM JudgeEvals
↗
Agents · DevTools

Terminal Coding Agent

Pip-installable CLI agent that plans edits via OpenAI/Anthropic, applies them through a diff manager, and executes inside E2B Firecracker microVMs — with guardrails that scan for leaked credentials (AWS, GitHub, Slack) before any edit is applied, covered by an automated test suite.

PythonE2B SandboxGuardrails
↗
Multi-Agent · AWS

CrisisCohort

Climate-resilience platform for smallholder farmers — a fleet of 10+ specialized AWS Bedrock agents (water optimization, climate migration, grant matching, disease diagnostics) on AWS Lambda, plus an on-chain water-rights trading contract in Solidity.

BedrockLambdaSolidity
↗
Agents · Voice

Voice Browser Agent

Talk to your browser — Deepgram streams speech to text, GPT-4o parses it into structured intents, and Stagehand executes them on real pages through Browserbase. Say "find me flights to Tokyo" and watch the browser do it.

DeepgramGPT-4oStagehand
↗
LLM Infra · Routing

LLM Router

Cost / latency / quality-aware LLM router on Next.js — NotDiamond picks the best model per request with a heuristic fallback chain across 5 providers, backed by continuous provider health checks.

Next.jsNotDiamond5 Providers
↗
Industrial CV · MLOps

NexusVision MLOps

Smart-factory visual inspection dashboard — Django REST + React TypeScript with live defect-detection simulation, a model registry tracking versions and metrics, and defect heatmaps across the production line.

Django RESTReact TSModel Registry
↗
Multimodal AI

Video Analysis System

Chat with any YouTube video — transcripts and OpenCV-extracted frames embedded for combined semantic + visual search, every answer cited with clickable timestamps. JWT auth, per-user rate limiting, an automated test suite, and Docker Compose with healthchecks.

GeminiFastAPIOpenCV
↗
RAG · Documents

Document Chatbot

FastAPI RAG chatbot — upload PDFs, ask questions, get cited answers. FAISS handles the vector search, the HuggingFace Inference API handles the generation.

FastAPIFAISSHuggingFace
↗
Healthcare · Accessibility

MediSign AI

Healthcare accessibility app in Flutter — communication tools for patients who can't speak or hear easily, a Gemini-powered medical chat, and telemedicine flows, all on a Firebase backend.

FlutterGeminiFirebase
↗
Urban AI · Kuala Lumpur

Cityscape AI

AI urban-planning app for Kuala Lumpur — graph neural network and reinforcement-learning traffic models behind a Flutter/Firebase frontend, with a Gemini RAG chatbot answering city-planning questions over local data.

GNN · RLFlutterGemini RAG
↗
HR Tech · DHL-Inspired

MomentumHR

Intelligent leave-logistics platform in Flutter + Firebase — separate HR and employee dashboards, bulk onboarding via Excel import, and a Gemini-powered chatbot that answers leave-policy questions so HR doesn't have to.

FlutterFirebaseGemini
↗
Semiconductors · Analytics

Chip Performance Platform

Semiconductor analytics platform — FastAPI backend simulating real-time chip telemetry with a Streamlit dashboard on top, and an ML model predicting chip failures before they happen.

FastAPIStreamlitML Prediction
↗
04b

Private & Client Builds

Eleven more shipped under NDA or private distribution — described by problem and outcome, since the code isn't mine to publish.

System Design · MLOps · 🔒 Private

MentorMind

ML-powered EdTech platform built as a full system-design showcase — nginx load-balancing two Django APIs, Postgres streaming read-replica with a custom DB router, Redis leaderboards, Celery queues, and a DVC-tracked MLOps loop with OpenCV proctoring, k6 load tests, and Kubernetes manifests with autoscaling.

DjangoKubernetesDVC MLOps
🔒
Data Engineering · 🔒 Private

CivicPulse

Civic-data pipeline pulling air quality (OpenAQ/WAQI), GTFS transit, and government notices into one schema — union-find dedupe across sources, weighted severity/proximity/recency ranking, and plain-language LLM summaries cached in Redis. Covered by an automated test suite in CI.

PythonNext.jsPostgres
🔒
Construction Tech · 🔒 Private

Qawel

WhatsApp-style construction project management on proper clean architecture (Riverpod, feature/domain layers) — project chat with voice notes, daily progress reports rendered to PDF server-side, and Google Cloud Translation across English, Arabic, and Urdu with full RTL support.

FlutterCloud FunctionsEN/AR/UR
🔒
Mental Health · 🔒 Private

Sushthotha

Bilingual (English/বাংলা) mental-wellness platform for Bangladeshi youth — a context-aware engine scores micro-interventions against mood, time of day, and screen time, alongside buddy circles and crisis-support flows. Multi-stage Docker deploy to Cloud Run.

React 19GeminiCloud Run
🔒
AgTech · 🔒 Private

Agritech Platform

Rice supply-chain platform for Bangladesh — five marketplace roles over a normalised Prisma schema with JWT-authenticated Socket.IO messaging, and an sklearn/XGBoost ensemble trained on district-level BORO/AMAN/AUS yield data feeding a full profit simulator.

XGBoostPrismaSocket.IO
🔒
.NET × React · 🔒 Private

Savora RMS

Restaurant management system on ASP.NET Core (.NET 10) + React 19 — POS with modifier-aware cart logic, kitchen display, inventory, and reservations across 12 pages. Every KPI (revenue, food/labor cost, channel mix) computed server-side from SQL Server, nothing hardcoded.

ASP.NET CoreReact 19SQL Server
🔒
Local LLM · RAG · 🔒 Private

ScholarMind

Academic research assistant that runs fully local — FAISS RAG over uploaded PDFs with Llama 3 via Ollama (Groq fallback chains), three-stage PDF extraction down to Tesseract OCR, a hand-rolled citation engine in four formats, plus AI flashcards and mind maps.

Llama 3FAISSRAG
🔒
Real-Time CV · 🔒 Private

Hack2Drive Controller

Autonomous racing controller for a Unity sim — a 5-task rate-monotonic schedule (5–25 ms periods) with deadline-miss instrumentation, custom OpenCV token detection with curve-aware lane geometry, and a YOLOv8 pipeline trained on an auto-labeled synthetic dataset.

OpenCVYOLOv8RTOS Model
🔒
HealthTech · v1.0 · 🔒 Private

VitalTrack Pro

Hospital-grade health monitor in Flutter + Laravel — Apple Health / Health Connect sync, vitals with 7-day trends, USDA + barcode food logging, RxNorm medication reminders, and branded PDF health reports. A separate doctor app gets live patient vitals with server-side threshold alerts.

FlutterLaravelHealth Sync
🔒
ERP · Laravel · 🔒 Private

School Management ERP

Production-ready school ERP on Laravel 13 + Filament 5 — the full chain works end-to-end: admission (with lottery) → attendance → invoicing → marks → tabulated results → report-card PDFs, plus payslips, ID cards, and a CMS-driven public website. Pest-tested, permission-driven roles throughout.

Laravel 13Filament 5Pest
🔒
Media · Bangla-First · 🔒 Private

সংবাদ News Portal

Prothom Alo-class Bangladesh news platform designed read-heavy and Bangla-first — Django 5 + DRF with Celery, Elasticsearch search, and Redis caching behind nginx, fronted by a Next.js 15 App Router UI. Full architecture rationale documented, Docker Compose throughout.

Django 5Next.js 15Elasticsearch
🔒
05

The Stack

Tools I reach for when something needs to exist by Friday.

// AI & ML

  • PyTorch · scikit-learn
  • XGBoost · LightGBM · Optuna
  • AWS Bedrock · Gemini
  • RAG — FAISS · pgvector
  • Local LLMs — Ollama · Llama 3
  • LLM judges · evals · SHAP

// Backend

  • FastAPI · Flask
  • Django + DRF · Celery
  • Node.js · Express · Nuxt.js
  • ASP.NET Core (.NET 10)
  • REST · WebSockets · Socket.IO
  • Prisma · SQLAlchemy · EF Core

// Frontend & Mobile

  • React · Next.js · React 19
  • React Native (Expo)
  • Flutter — Riverpod · Provider
  • TypeScript
  • Tailwind CSS · Three.js
  • i18n — EN · BN · AR · UR · MS

// Cloud & DevOps

  • AWS — Lambda · Bedrock · DynamoDB · Cognito
  • GCP — Cloud Run · Firebase
  • Docker · Kubernetes
  • GitHub Actions CI/CD
  • Redis · Prometheus · Grafana
  • PostgreSQL · Supabase · MySQL · SQL Server · Firestore
06

Receipts

Education, certifications, and the program that started it all.

Education

Universiti Teknologi Malaysia

Bachelor of Software Engineering — Johor Bahru. Graduated July 2026, First Class.

Certification · Aug 2026

AMD — ROCm Certified Associate

GPU programming with ROCm/HIP. Scored 100% on all five course quizzes and the final exam, first attempt. Valid to Aug 2028.

Certification · Mar 2025

AWS — SageMaker, LLMs & Prompt Engineering

Amazon SageMaker, Large Language Models, and Prompt Engineering certifications.

Certification · Aug 2024

IBM — Full Stack Software Developer

Professional Certificate covering full-stack development, cloud native practices, and containerization.

Certification · Jun 2024

Microsoft Azure AI

Azure OpenAI Service, Natural Language Processing, and Document Intelligence.

Achievement

Headstarter SWE Resident

Selected for a competitive 6-month US-based engineering residency — shipped ML and full-stack projects end to end.