AI Engineer & Full-Stack Architect • Christchurch, NZ

Transforming Brittle AI Loops Into Deterministic Enterprise Workflows.

Modern generative AI promises autonomy, but raw LLMs deployed into production quickly degrade into token burn, context drift, and non-deterministic failure loops. I architect observable, cost-effective full-stack platforms and state-machine-driven agent systems that deliver deterministic outcomes at scale.

Explore Architecture Deep Dives

Agentic Coding, Tool Use & Automated Workflows 9+ Years Production Engineering • Master of AI (UC)

The Industry Bottleneck

Uncontrolled Agent Autonomy

× Unbounded Token Burn: Unchecked prompts and massive history dumps inflate inference budgets without improving reasoning.
× Context Drift & Hallucination: Agents forget core constraints after multi-turn tool calling, injecting subtle bugs into codebases.
× Brittle Feedback Loops: Agents hallucinating CLI outputs or spinning in circles without deterministic exit criteria.

My Architectural Approach

Observable & Deterministic Systems

✓ Strict State Machine Boundaries: Deterministic transitions, typed schema gates (Zod), and sandboxed execution environments.
✓ Context Token Optimization: Surgical file-level slicing, smart `.aiignore` filters, and aggressive prompt caching.
✓ Automated TDD & Verifiable CI/CD: Self-healing regression test suites and containerized delivery pipelines (Docker + Caddy).

Mainuddin Talukdar

Christchurch, New Zealand

GitHub LinkedIn

About Me & Applied AI Focus

Bridging Agentic AI Workflows & Production Full-Stack Systems.

I am a software engineer and Master of Artificial Intelligence student at the University of Canterbury (Christchurch, NZ) with nine years of experience designing, building, and shipping production systems — most of it in Java and Spring Boot — now focused on applied AI with Python.

The backend, object-oriented, and distributed-systems foundations I built across nine years in Java transfer directly to Python, which I use across AI projects, data pipelines, and microservices. I build LLM-powered agents end to end, using Claude and Claude Code, Cursor, and Codex well beyond chat — for agentic coding, tool use, and automated workflows — alongside full-stack delivery in TypeScript and Next.js.

Master of Artificial Intelligence

University of Canterbury • Current (NZQA Level 9)

IEEE Publication (ICCIT)

“A Stacked Meta-Model Framework for Diabetes Prediction”

MS in Computer Science

Maharishi International University, USA (2025)

BS in Software Engineering

University of Dhaka, Bangladesh (2016) • DhakaThon Champion

Verified Career Credentials & Full CV

Direct phone and email contact details are gated behind an automated intent verification system to eliminate scrapers and spam.

Guiding Principles & Operating System

Mission & Engineering Decision Matrices

Great software is not defined by complexity, but by the clarity of trade-offs made during its conception.

Core Vision Statement

“Making the world of Agentic AI usable, maintainable, and extensible.”

Transforming unpredictable probabilistic AI experiments into rock-solid, production-grade business assets.

Customer Obsession

Technology serves human outcomes. I measure architectural success not by lines of code or complex jargon, but by end-user delight, operational resilience, and direct business ROI.

Sustainable Architecture

Minimalist dependencies, zero-bloat JavaScript footprints, strict schema boundaries, and deterministic state machines. Systems built to outlive technology hype cycles.

Trade-Off Decision Matrix

Every engineering choice balances Cost, Latency, and Maintainability. I eliminate premature optimization by deploying quantitative telemetry and cost-budget thresholds.

Architectural Trade-Off Matrix

Cost vs. Latency vs. Maintainability

Dimension	Naive / Hype Pattern	Engineered Solution	Key Metric Impact
Agent Context & Memory	Dumping entire repo & history in raw prompts	Surgical `.aiignore` + AST context slicing + prompt cache	-70% Token Costs
Web Rendering Model	Heavy SPA with 2MB JavaScript bundle	Astro SSR with zero JS default + React Islands	< 50ms TTFB / Zero CLS
Lead Qualification	Exposed mailto links & unauthenticated scraping	AI Intent Gate + Supabase RLS + Resend Automation	Zero Spam / 100% Verified

Verified Technical Core

Skills & Engineering Matrix

Categorized breakdown of technical competencies, tools, and methodologies with verified production recency and quantifiable impact.

AIActive Production (2024–2026)

AI Agents, LLMs & Agentic Coding

Claude and Claude Code, Cursor, Codex used for agentic coding, tool use, and automated workflows; LLM integration, agent architectures, prompting strategies, RAG, benchmarking, and evaluation for real SME operations.

Keywords & Tools

Claude CodeCursorCodexAgent ArchitecturesTool CallingRAGPrompt StrategiesLLM Benchmarking

BACKEND9 Years Core Production

Backend Engineering (Java → Python)

Nine years in Java and Spring Boot — REST APIs, microservices, OOP and design patterns, SQL (PostgreSQL, Oracle) and NoSQL — with the exact same distributed-systems fundamentals now applied in Python for services, data pipelines, ML, and automation. Java experience transfers to Python directly.

Keywords & Tools

Java 11/17/21Spring BootPythonMicroservicesREST APIsRedis CachingPostgreSQL / OracleKafkaDistributed Systems

FULLSTACK9 Years Production

Full-Stack Web & Mobile Ecosystems

TypeScript/JavaScript, Next.js, React, React Native (shipped iOS & Android apps to App Store/Play Store), Node.js, Astro, REST API design, Supabase, PostgreSQL, and Vercel deployment with accessible functional component design.

Keywords & Tools

TypeScriptNext.jsReactReact Native (iOS/Android)AstroNode.jsREST APIsSupabase / Postgres

BACKENDMaster of AI (Level 9)

Python, Data Pipelines & Applied ML

Python for autonomous agent services, ETL data pipelines, machine learning, and deep learning. NumPy, Pandas, PyTorch workflows, and Spark/Hadoop processing over multi-terabyte datasets, backed by deep CS and enterprise architecture foundations.

Keywords & Tools

PythonNumPy / PandasPyTorchSparkSQLHadoopData PipelinesMeta-Learning MLDeep Learning

CLOUD9 Years Production

Cloud Infrastructure, AWS & CI/CD

AWS (Lambda, DynamoDB, S3, NoSQL), Docker containerization, Redis caching layers (cutting read latency >80%), GitHub Actions CI/CD pipelines, Caddy reverse proxy, and operational telemetry dashboards.

Keywords & Tools

AWS (Lambda, S3, DynamoDB)DockerRedis CachingCI/CD PipelinesGitHub ActionsCaddy Reverse ProxyKafkaTDD / Vitest

LEADERSHIPContinuous Practice

Ways of Working & Client Collaboration

Self-directed, end-to-end ownership of features from whiteboard to deployment. Cross-functional leadership, client-facing communication, technical trade-off analysis, and clear written and verbal collaboration.

Keywords & Tools

End-to-End OwnershipClient-Facing DeliveryTechnical LeadershipTrade-Off AnalysisAgile / JiraCode Reviews

Career Trajectory & Track Record

Experience & Quantifiable Impact

Nine years of engineering leadership, production cloud systems, and measurable performance optimizations across enterprise platforms.

Aug 2024 – Feb 2026 New Jersey, USA

Software Development Engineer II

Aperia Solutions (Client: Fiserv)

Impact: Optimised Java/Spring Boot APIs by ~40% under peak load; led Java 11 to 21 upgrade and cut read latency >80% with Redis for millions of daily users.

• Provided technical leadership on assigned features, owning requirements analysis, design, build, test, deployment, and documentation across payment and reporting services.
• Built and optimised RESTful APIs and enterprise integrations in Java and Spring Boot, improving response time by ~40% under peak load; led the Java 11 to 21 upgrade and integrated Redis caching to cut average read latency by over 80% for real-time updates serving millions of daily users.
• Automated build, test, and deployment through CI/CD pipelines with Git workflows, Jira tracking, and monitoring, communicating technical trade-offs clearly to internal and client stakeholders.

Jun 2019 – Aug 2023 Pennsylvania, USA (Remote)

Software Developer → Senior Software Developer

Vantage Labs

Impact: Shipped Java/Spring Boot services, React frontends & 2 React Native apps to App Store/Play Store; led Python pipelines lifting revenue ~9.7%.

• Built full-stack products across Java and Spring Boot services and TypeScript/React front ends — including two React Native apps shipped to the Apple App Store and Google Play — with REST APIs, Firebase authentication, and real-time data.
• Led data-analysis pipelines in Python with Hadoop and Spark to process large-scale performance data, driving athlete recommendations that lifted revenue ~9.7% annually.
• Wrote unit and integration tests, resolved UI performance issues, deployed to AWS (Lambda, S3, NoSQL) with monitoring, and was promoted to Senior for sustained impact.

Jun 2016 – Jun 2019 Dhaka, Bangladesh

Software Developer

TigerIT Bangladesh Ltd

Impact: Delivered national-scale government systems serving 100M+ citizens built on Java, Spring Boot & Oracle, cutting licence processing by ~30%.

• Delivered national-scale government systems serving over 100 million people, built on Java, Spring Boot, Oracle, and WebLogic, cutting licence processing time by ~30%.
• Built automated test suites and reporting pipelines with SparkSQL and Kafka to improve reliability and deliver real-time operational insight over 35 TB of data.

Shipped Systems & Engineering Depth

Featured Projects & Engineering Retrospectives

Real production applications with live demos, transparent architecture notes, and lessons learned through engineering adaptations.

Conversational AI & Proximity Matching Live on Subdomain / Beta

Live App

TradiePulse Conversational AI Agent

Stack: Next.js Frontend + Python Agent Service (NLP Matching) + Supabase + Docker + Caddy

A conversational AI agent that lets a customer describe a household problem in plain language, the way they would to a person. The agent interprets the request, then finds the nearest available and suitably qualified tradesperson — plumber, electrician or mechanic — and connects the two. Built with a Next.js front end, a Python agent service for language understanding and matching, and Supabase for tradie profiles, availability and jobs; containerised with Docker and served on Hostinger VPS behind Caddy.

Measured Delivery: Interprets customer household problems in plain language, matches nearest qualified tradesperson (plumber, electrician, mechanic), and connects both parties.

Engineering Retrospective: Challenge & Solution

The Engineering Challenge:

Directly passing unconstrained conversational text to tradespeople caused ambiguity in job scopes and dispatch errors.

The Architectural Adaptation:

Architected a structured entity extraction pipeline in the Python agent service that translates ambiguous natural language descriptions into standardized trade categories, urgency levels, and location parameters before database matching.

Next.js Python Agent Supabase PostgreSQL Docker Caddy Tailwind CSS

Adaptive Learning & Python ML Service Production / Live Demo

Live App

MathQuest Adaptive Learning App

Stack: Next.js + Supabase + Vercel + Python Scoring Microservice

An interactive maths learning app for primary-school children, built on Next.js, Supabase and Vercel. Children work through puzzle-based lessons while the app tracks how they engage — time per question, attempts and accuracy — in Supabase, then scores each child and suggests what to practise next. The scoring and suggestions run as a separate Python microservice, proving the end-to-end workflow ahead of a fully trained model.

Measured Delivery: Tracks fine-grained engagement (time per question, attempts, accuracy), scores student mastery, and dynamically suggests practice exercises.

Engineering Retrospective: Challenge & Solution

The Engineering Challenge:

Embedding scoring heuristics directly in frontend components created inconsistent evaluation metrics and blocked future ML model integration.

The Architectural Adaptation:

Decoupled telemetry collection from recommendation logic into a standalone Python microservice, allowing independent scoring algorithms and smooth migration path toward deep reinforcement learning models.

Next.js Python Microservice Supabase Vercel Tailwind CSS Web Audio API

Infrastructure & Delivery Harness Active Production Standard

Live App

Dockerized Microservices & CI/CD Pipeline

Stack: Docker Multi-Stage + GitHub Actions CI/CD + Caddy Reverse Proxy + Supabase RLS

Each project runs as its own Dockerised service, ships through a strict GitHub CI/CD pipeline with automated tests and smoke checks, and is served on its own subdomain via Caddy — establishing an observable, repeatable path from commit to production.

Measured Delivery: Automated test execution, smoke checks, and zero-downtime subdomain deployments serving live apps on mainuddintalukdar.cloud.

Engineering Retrospective: Challenge & Solution

The Engineering Challenge:

Manual deployment and ad-hoc subdomain routing caused configuration drift and intermittent SSL renewal failures across multiple microservices.

The Architectural Adaptation:

Automated the deployment pipeline using GitHub Actions with SSH orchestration and Caddy automated ACME certificate management, guaranteeing deterministic releases.

GitHub Actions Docker Caddy Reverse Proxy TypeScript Vitest Zod Gates

Academic Research & Achievements

Peer-Reviewed Publication & National Hackathon

IEEE Publication • ICCIT Peer-Reviewed

“A Stacked Meta-Model Framework for Diabetes Prediction”

Applied advanced feature engineering, ensemble stacking, and meta-learning algorithms to a real-world biomedical machine learning dataset to achieve superior predictive accuracy and generalization.

View IEEE Xplore Paper

Champion • 1st Place 100+ Teams

DhakaThon 2015 National Hackathon

Awarded 1st Place among 100+ competing engineering teams in a national two-day rapid application development hackathon, demonstrating rapid prototyping, architectural execution, and effective presentation under tight constraints.

National Engineering Championship