AM
Available Resume

Open to Work — AI Quality & Evals Engineer · Toronto / Remote · No sponsorship needed

Aiden Mak

Software Engineer — AI Quality & Evals

Anyone can demo AI. I ship it to production.

I build agentic AI products, then make them trustworthy — eval harnesses, 1,900+ automated tests, LLM-judge evals, and production tracing that show they hold up for real users.

See the work Download resume
Read my story →
  • AI agents
  • Evaluation
  • Full-stack products
  • RAG systems
  • UAT / QA
Portrait of Aiden Mak
Toronto — where DecisionEase and the HR Intelligence Platform get built.
Building

Agent workflows, evaluation harnesses, and AI systems that expose evidence.

Learning

Healthcare AI, speech-to-text testing, product reliability, and user trust.

Open to

AI product engineer, AI consultant, founding engineer, and practical AI product teams.

1,909 automated tests on one project — proof over promises
14 specialized AI agents across two live systems
97% faster API — 6.8s to 0.2s after my RAG optimization in Human Rights RAG
LIVE DecisionEase is running right now — open it →
Work

Selected work

The HR Intelligence Platform and DecisionEase are the two flagships; the rest show range across RAG, review intelligence, UAT, education, and data systems.

Flagship 01

HR Intelligence Platform

Enterprise-style HR agent platform with 8 specialist agents, RAG, FastMCP, a BambooHR connector, PII/RBAC/security layers, and observability.

Problem
HR work spreads across policy, leave, benefits, compliance, analytics, and employee data.
My role
Independent AI/full-stack builder across agent architecture, backend, dashboard, deployment, and tests.
Built
8 LangGraph agents, RAG knowledge layer, Flask API, FastMCP server, observability, RBAC, and PII masking.
Evidence
28 MCP tools, 8 resources, 5 prompts, 1,909 tests, GCP Cloud Run, Prometheus, Grafana, LangSmith.
Finding
Agents become useful when routing, tools, retrieval, security, and tests are visible system parts.
LangGraph Flask MCP ChromaDB
Full case study Live dashboard GitHub
github.com/aidenmak0624/HR-Intelligence-platform
HR Intelligence Platform demo
Live demo capture from the repository Verified → repo
Flagship 02

DecisionEase

AI decision-support PWA built around a 6-agent backend, 4-layer memory, FastAPI, Next.js, PostgreSQL / pgvector, Redis, and a product loop from capture to reflection.

Problem
Personal decisions scatter across notes, tasks, calendars, moods, goals, and memory.
My role
Creator and full-stack builder for the agent backend, PWA surfaces, memory architecture, and product loop.
Built
6-agent decision-support flow, 4-layer memory, FastAPI backend, Next.js PWA, pgvector, Redis, and integrations.
Evidence
27 API routes, 32 migrations, 7 internal MCP tools, 3 external tools, 5,000 ms agent-to-agent (A2A) timeout, and backend tests.
Finding
Decision support needs a loop: capture, compress, review, decide, schedule, and reflect.
FastAPI Next.js pgvector Redis
Full case study Live product Landing page
decisionease.vercel.app
DecisionEase app — Today view
The live app: capture, plan, decide, reflect Live → open
More cases
01

ReviewLensAI

Problem
Teams need to turn messy product reviews into grounded insight without losing source evidence.
My role
Full-stack AI builder across ingestion, RAG chat, citation UX, edge functions, and evaluation.
Built
React/Vite app, Supabase Edge Functions, Pinecone retrieval, GPT-4o chat/vision, and 3-worker insight pipeline.
Evidence
4 ingestion modes, 5 edge functions, top-K 8 retrieval. 168/173 Vitest passing — the 5 failures documented, not hidden · 6/6 promptfoo LLM evals.
Finding
AI answers feel trustworthy when claims can open the exact review that supported them.
React Supabase Pinecone RAG
Live product GitHub
02

Zokforce AI Consultancy

Problem
Clinical AI has to work through noisy audio, multi-speaker speech, accents, code-switching, and trust constraints.
My role
AI Engineering Intern — volunteer Jan – May 2026, intern May 2026 – present — clinical simulation tools, STT accuracy testing (WER/CER), bilingual UAT plans, and defect verification.
Built
clinsim and stt_uat tools for labeled multi-speaker clinical audio, configurable TTS, normalization, and scoring.
Evidence
Python, numpy, ffmpeg, pytest suites, bilingual UAT plans, defect register, and Cantonese/code-switched test flows.
Finding
Acceptance testing is where model quality meets language, workflow, acoustic reality, and release risk.
UAT Python STT pytest
Company
03

GenAI Customer-Support Assessment

Problem
Support questions span structured order data and unstructured policy PDFs — one assistant has to know when to query a database, when to read documents, and when to do both.
My role
Solo build for a timed GenAI take-home assessment — architecture, agents, retrieval, tests, and deployment.
Built
LangGraph supervisor with four explicit routes (SQL, RAG, hybrid with a synthesizer node, general), ChromaDB retrieval with lexical reranking, read-only SQL validation, and the same backend exposed as a Streamlit chat app and a FastMCP server.
Evidence
Public repo, Playwright E2E plus pytest, and a citation (file/page/line or table/SQL) on every answer.
Finding
Routing is a product decision — four explicit routes made the system's behavior predictable and auditable.
LangGraph RAG ChromaDB FastMCP
GitHub
Additional works

Earlier projects that shaped the same pattern.

Lighter entries that show the range behind the main case studies: ordering platforms, hackathon builds, education tools, team apps, and civic data.

FieldInvoice AI

Hackathon build: a messy field note becomes a verified invoice and an instant Stripe payment link with QR — LLM extraction with a deterministic fallback, documented with PRD, BDD, and ADRs for all 19 modules.

Next.js Stripe LLM

Food Toxicity Report App

Team of three — lead contributor. Food-safety reports from a barcode or product name: Open Food Facts and openFDA signals, risk scoring, AI analysis, JWT auth, and a Playwright E2E suite.

Next.js openFDA JWT
GitHub

CMHR Upstander Program

Flask and Gemini learning journey for the Canadian Museum for Human Rights, guiding users through a 6-stage human-rights education flow.

Flask Gemini EdTech
GitHub

Human Rights RAG

Topic-scoped RAG platform over 25+ UN/OHCHR documents with 9 ChromaDB collections, citations, and difficulty-aware answers — plus a 97% faster API (6.8s → 0.2s) after optimization.

RAG ChromaDB Gemini
GitHub

TeachReach

Java Android tutoring marketplace with tutor search, scheduling, booking, payment validation, messaging, HSQLDB, and layered tests.

Java Android HSQLDB
GitHub

311 Service Request Analysis

Python civic-data analysis over 16 years and 1.6M+ Winnipeg 311 records using FP-Growth, STL, MAD, and stakeholder reporting.

Python Data mining Civic tech
GitHub
Findings

Short notes on AI systems, evaluation, and product judgment.

Observations from building, testing, and learning — written for people who want to understand how I think, not just what I shipped.

EssayEvaluation · 5 min read

Most of my RAG evals didn't need an LLM judge

Six promptfoo cases grade the live ReviewLens AI pipeline 6/6 — a checkable citation contract made most grading deterministic, and showed where a judge is still missing.

Read essay
LinkedInAI Systems · Apr 2026

What I learned directing AI agents for 6 months as a recent grad

Agents fail through overreach and under-finish — harness engineering, not bigger prompts, is what gets AI-built work to done.

Read on LinkedIn
EssayAI Systems · 4 min read

AI Needs Evidence, Not Confidence

Good AI work gives people a way to inspect the source, confidence, trace, or failure mode behind a response.

Read essay
EssayEvaluation · 5 min read

Why 1,909 tests mattered more than the demo

The demo showed the idea. The tests showed where the product could break before a user touched the real path.

Read essay
EssayAI Systems · 6 min read

Why tool-use needs observability

When an agent calls a tool, the user experiences the result. Tool-use needs traces, timeouts, and a visible failure path.

Read essay
EssayAI Systems · 4 min read

Agents Need Boundaries, Memory, and Tests

Multi-agent systems become useful when routing, tools, memory, retrieval, and fallback behavior are explicit.

Read essay
NoteEvaluation · 2 min read

Evaluation is product work

The hardest tests are not always technical. They come from what users need to trust, repeat, and act on.

Read note
NoteProduct Lessons · 2 min read

Real users reveal edge cases faster than demos

Demos reward the happy path. UAT teaches you where language, workflow, timing, and expectation diverge.

Read note
NoteAI Systems · 2 min read

Healthcare AI is a trust problem before it is a model problem

Accents, noise, clinical language, code-switching, and data residency all shape whether the system is usable.

Read note
Story

Building where real users make systems honest.

I started in computer science because I liked turning complexity into working software. I stayed with AI because the hardest part is rarely the model alone. It is the messy workflow around it: data, trust, testing, product constraints, and the people who need the system to work.

2020 - 2025

University of Manitoba

BSc Honours Computer Science, grounding my work in software engineering, AI, databases, and algorithms.

2024 - 2025

Full-stack and AI product work

Built systems across RAG, product dashboards, multi-agent workflows, and user-facing AI tools.

2026

Zokforce AI consultancy

UAT/QA work on clinical encounter simulation, STT evaluation, bilingual test plans, and defect tracking.

Now

Practical AI systems

Focused on AI that can be tested, explained, monitored, and improved after real users touch it.

How I think about building

Messy requirements are signal.

When a problem feels unclear, I look for the user workflow, failure modes, and evidence needed for a good decision.

Evaluation is product work.

AI quality is not just passing tests. It is knowing what behavior matters, how it fails, and who needs to trust it.

Agents need boundaries.

Multi-agent systems become useful when routing, tools, memory, retrieval, and fallback behavior are explicit.

Under-claim, then prove.

I would rather show the test, the trace, the citation, or the defect log than hide behind impressive language.

Now

What I am focused on right now

Updated July 9, 2026. Status: Open to Work.

Current focus

Building practical AI systems where outputs are testable, cited, traced, and useful to real workflows.

Learning

Healthcare AI testing, STT robustness, agent memory, product reliability, and evaluation methodology.

Building

DecisionEase iterations, case-study writing, UAT frameworks, and clearer notes from past projects.

Open to

AI engineering roles in Canada, applied AI teams, product engineering teams, and thoughtful collaborations.

Resume

Resume at a glance

AI / Full-Stack Engineer with hands-on work across multi-agent systems, RAG, evaluation, testing, and product software. Authorized to work in Canada through May 2029 — no sponsorship required; permanent residence in progress.

Download PDF

Experience

  • Zokforce - AI Engineering Intern, AI Consultancy
  • Independent AI / full-stack systems builder
  • University of Manitoba - BSc Honours Computer Science

Technical skills

AI / ML

Multi-Agent Systems, LangGraph, LangChain, MCP, RAG Pipelines, Prompt Engineering, OpenAI API, Gemini, ChromaDB, Pinecone, Sentence Transformers

Programming

Python, TypeScript, JavaScript, Java, C/C++, SQL

Full-stack

Flask, FastAPI, Next.js, React, Node.js, PostgreSQL, Redis, SQLAlchemy, REST APIs, WebSocket, Stripe, Tailwind CSS

Infrastructure

Docker, GCP Cloud Run, CI/CD, Git, Linux, Prometheus, Grafana, Playwright, pytest

Practices

Agile / Scrum, TDD, SOLID, OOD, Code Review, UAT, QA, AI evaluation

Languages

Cantonese (native), Mandarin (fluent), English (fluent)

Contact

Let's talk.

I'm open to AI quality & evals, QA automation, and AI engineering roles in Toronto — remote works too. If any of this sounds like a problem you're hiring for, my inbox is open.

Authorized to work for any Canadian employer through May 2029 — no sponsorship required.

LinkedIn
GitHub mcwaiden000@gmail.com