AI engineer in Ho Chi Minh City, building practical systems with LLMs, agents, OCR and automation — mostly for people who are learning something hard.
Agents · Edtech · OCR
Now — AI Engineer at GrowtricsAI
Download CV ↓
> Fig.00 — one study night, four moments
01 — A real moment
She did the work and still can't see the mistake. Quiz is Thursday at 8. This is the moment I care about building for.
02 — Listening first
The agent doesn't paste the solution. It asks which part feels shaky, keeps what it already knows — this week's topic, the quiz — and breaks the fix into three small steps.
03 — Reading the page
OCR is one capability here, not the point: a phone photo of handwriting becomes text the agent can point at — this line, this step.
04 — The next step
The win isn't a perfect answer from a machine. It's a student who can do #4 on her own — and then goes to sleep.
> Fig.01 — the parsing problem
Messy input
A phone photo, not a clean PDF.
Skewed page, a coffee ring, a margin scribble, a printed diagram and a running header — the kind of input real students actually send.
Crop & segment
Find the regions that carry meaning.
Layout detection separates body text, the figure and the handwritten note, then crops each region upright so nothing is lost to the noise around it.
Parse & clean
Turn pixels into structure.
OCR reads the text, broken lines are rejoined, page furniture is dropped, the equation stays an equation and the figure gets a caption it can be searched by.
Chunk with provenance
Small pieces that remember where they came from.
Text and figure become self-contained chunks, each tagged with section, page and links to its neighbours — so “see Fig. 3.2” still points somewhere.
Embed
Meaning becomes a position.
Every chunk — text and image description alike — is embedded into a vector space where related ideas land near each other and can be found by meaning.
Retrieve & answer
The agent answers from the page.
A student's question pulls the nearest chunks, the agent packs them into context and answers — citing p.42 and Fig. 3.2 instead of guessing.
Visual explanation — not a live tool. Nothing is uploaded or processed here.
> Intro
I'm Michael Le, an AI engineer based in Ho Chi Minh City. I build systems that meet people where they are — turning questions, notes and messy documents into a useful next step.
Thread 01
Autonomous systems and practical agent workflows.
Thread 02
Extracting structure from unstructured content.
Thread 03
Better tools for learning and growth.
Raw text, documents and photos in; clean, structured context out — so an agent retrieves the right passage or figure and answers from it, with sources.
Tools for personalised learning experiences with AI-driven content recommendations.
The area of my work at GMO-Z.com RUNSYSTEM: reading documents across scripts and layouts.
> Writing & Research
> Life — off the clock
Most of the quality in an agent's answer is decided before the model ever runs. This is how I approach turning unstructured, multimodal material into context an LLM can use — seven stages, each with an output you can inspect.
Traced example One phone photo of a physics textbook page — p.42, body text, a printed trajectory diagram, a margin note and a coffee ring — followed through all seven stages to one student question.
Ingest
PDFs, scans, screenshots and phone photos arrive in different formats and quality. Each item is typed, split into pages and given a source record, so every later chunk can point back to where it came from.
IMG_2231.jpg · 3024×4032tilted · shadow · coffee ringno text layer
src: phys-ch3 · p.42type: photo / textbook pagerotated 4° · deskewed
Parse & clean text
Text is read (OCR where there is no text layer), then normalised: hyphenated lines rejoined, running headers and page numbers dropped, equations and lists kept as structure rather than flattened prose.
CHAPTER 3 · MOTION 42At the top of its path the ver-tical velocity v_y = 0 . (see Fig
§3.4 Projectile motionAt the top of its path, verticalvelocity v_y = 0 (see Fig. 3.2).
Crop & read images
The page is segmented into regions — body text, figure, handwriting. The diagram is cropped tight to its bounds, then described and labelled so an agent knows what it shows and which paragraph cites it.
whole page · 1 imagediagram + margin scribble+ shadow, all mixed
fig-3.2 · crop 612×388“parabola; v_x constant, v_y → 0 at apex”
Structure
Clean text and image regions merge into chunks with a title, a place in the hierarchy, a link from the figure to the paragraph that cites it, and metadata the agent can filter on.
loose text blocksa separate image cropno link between them
chunk §3.4#2 · kind: text cites → fig-3.2 topic: projectile · ch 3
Embed
Each chunk, and each figure description, is encoded as an embedding — a vector of its meaning — and stored with its metadata in an index built for semantic search.
§3.4#2 (text)fig-3.2 (description)
[0.12, −0.48, 0.07, …][0.09, −0.51, 0.11, …]+ metadata · src p.42
Connect the agent
Retrieval is wired in as a tool. The agent decides when to search, rewrites the question into a better query and chooses filters, combining what it finds with its own plan and memory.
“why is it 0 at the top??”
search(“vertical velocity at apex of projectile”, filter: ch 3)
Retrieve & respond
Matching chunks are retrieved and re-ranked, packed into the prompt with their sources, and the model answers from them — pointing at the page, figure or line it used.
top hits: §3.4#2 · fig-3.2dropped: ch 5 (off-topic)
“At the apex the ball stops rising before it falls, so v_y = 0.” — p.42, Fig. 3.2
Explanatory workflow from my work — not a live tool on this site. Nothing is uploaded, processed or stored here.