Profile · Desoul

Rendered from this record. Re-export the PDF after a render.

~/Documents/cv/Nguyen-Bao-Le-CV.pdf Open full size ↗ zoom

Professional summary

Data analyst with seven years across Business Intelligence, Data Analytics and applied AI, and a finance background. I build the full path from raw data to decision: ETL pipelines and warehouse models, the dashboards business teams run on, and the machine-learning and generative-AI systems that turn unstructured market and player data into answers. I also design and ship production LLM systems end to end — retrieval, grounding gates, governed agent execution and automated evaluation — and I decide with measurements rather than vendor claims: I raised retrieval accuracy from 79% → 93% on a 1,253-document corpus, and I have killed features I had already built when the measurement said they would cost recall. Domain depth in mobile game publishing for the SEA market: user acquisition, monetization, retention and competitive benchmarking. Available now for contract work, full time.

VNG dates

March 2019 – May 2026

Rendered from this record and live at nguyenbaole.pages.dev — Cloudflare Pages, free tier, unmetered bandwidth, deployed by `wrangler pages deploy public --project-name nguyenbaole`. The videos are served from Pages itself, which is what keeps them inside Cloudflare's video policy; proxying them from elsewhere would not be.

~/Documents/portfolio/public/index.html Open full size ↗ zoom

Availability line

Available now, full time. Remote from Ho Chi Minh City — a full working day with Asia and Australia, and the whole European morning. Scoped builds, fixed-price audits, ongoing by the month.

Run-cost paragraph

Run cost is a design constraint, not an afterthought. Everything here runs on a subscription you already pay for or on models running locally — not a per-token bill that grows with your success. That is a measurement habit rather than a preference: benchmark the free option against the paid one and pick on the result. 8.8% word error rate, chosen voice. 79% → 93% retrieval accuracy, measured.

Title · 67/70

AI Engineer | RAG, LLM Evaluation & Agents | 7 Years Data Analytics

Overview — visible before the fold · 242/250

I build LLM systems you can trust in production: retrieval that finds the right passage, grounding gates that refuse unsupported claims, and evaluation that catches a bad prompt change before your customers do. 79% to 93% retrieval, measured.

Skills (20/20)

Python, Large Language Model (LLM), RAG, Prompt Engineering, AI Agent Development, AI Model Evaluation, FastAPI, SQL, Data Engineering, ETL Pipeline, Apache Airflow, Apache Spark, PostgreSQL, DuckDB, Machine Learning, Data Analysis, Business Intelligence, Tableau, Looker Studio, Microsoft Power BI

Headline · 72/220

Data & Applied AI · LLM systems with retrieval, grounding and evaluation

About

I build LLM systems you can trust in production: retrieval that finds the right passage, grounding gates that refuse unsupported claims, and evaluation that catches a bad prompt change before your customers do. 79% → 93% retrieval, measured.

What I get hired to do:

Make an LLM system trustworthy. Retrieval that finds the right passage, a grounding gate that refuses claims the source does not support, and an evaluation harness that tells you when a prompt change made things worse — instead of a customer telling you. (79% → 93% retrieval accuracy, measured · 3,600 decisions in one evaluation)

Put agents to work safely. Coding and ops agents running real tasks under written rules: forbidden actions stop and ask a human, work happens on an isolated branch, and nothing lands without a reviewed diff. (8 active projects governed · 117 test modules, TDD)

Turn data into decisions. Seven years of it: pipelines over terabytes a day, dashboards people actually open, and churn / LTV / segmentation models pointed at a number someone is accountable for. (−80% average cost per install · +30% monthly game revenue)

Available now for contract work, full time. Remote from Ho Chi Minh City — a full working day with Asia and Australia, and the whole European morning.

No separate profile to write: upload the CV, then a ~20-minute one-way AI interview. The CV above IS the profile here, and the prep below is for reading it back.

What the interview is

~20 minutes, voice or voice+video, recorded, no human on the other end.
6 to 8 questions, each with a follow-up generated from what you just said.
The first 2–3 minutes verify the CV: "walk me through the project you listed". It is reading the document, so every claim in it needs a story behind it.
Difficulty ADAPTS. A strong answer earns a harder question; a vague one earns a broader, easier question — and that caps the score you can still reach. An early weak answer costs more than a late one.
Outcome is a tier — Declined, Qualified, Expert, Verified Expert — not a per-question mark.

What it scores

specificity — a number, a name, a date. Not "significant improvement".
ownership — what YOU decided, not what the team did.
technical depth — the follow-up is the real question; the first one is the setup.
communication compression — how much you say per second. 90 seconds to 2 minutes per answer. Over 3 and it needs cutting; under 90 and it reads as thin.

Rehearse these 9

Walk me through a time you made a decision the data disagreed with you on.

The keyword pre-filter. You had already built it. 3,600 labelled decisions said it would cost recall, so you deleted it. Ownership plus specificity in one story.

Tell me about a bug that was hard to find.

Four of five ingestion sources emitting a fabricated popularity metric the model then reasoned from — a repository with 38,586 stars scored as if it had none. Found by checking output against the real API rather than trusting it.

How do you know your retrieval is any good?

A hand-labelled set, and the number moved 79% → 93% on 1,253 documents. Name the method before the result: keyword vs hybrid keyword-plus-embedding.

Tell me about your role at VNG.

Split it the way the CV does — the analyst track and the AI track, seven years, same job. Close on −80% cost per install. Do not narrate the org chart.

Do you use AI to write your code?

Yes, and say so first. Then move immediately to what you do with the output — the fabricated metric, the killed feature. The question is a proxy for "can you fix this when it breaks", so answer that one.

Walk me through the data pipeline you owned at VNG.

The scale IS the specificity: terabytes a day, warehouse models behind 30+ published titles. Name the stack in the order the data moved through it, not as a tool list.

You supply: One failure and what you changed after it — a late-arriving source, a schema change that broke a dashboard, a job that half-ran and reported success. Depth probes go at failure handling, never at the happy path.

How did you validate the churn model?

Name the baseline before the model. The work aimed at targeted marketing spend, so the answer ends at a marketing decision, not at an accuracy score.

You supply: Features, train/test split, and what the baseline was. "Better than guessing" is the answer that caps the score you can still reach.

How did you size an A/B test?

A/B testing is on the CV with nothing behind it yet. Expect the follow-up to be about what you did when a result came back inconclusive.

You supply: One real test: the metric, the minimum effect worth detecting, how long you ran it, and the call you made at the end.

What made your dashboards ones people actually used?

Three BI tools on one CV invites "why three?" — the honest answer is that different teams had standardised differently, and it takes two sentences. Then move to the Sankey and waterfall journey work, which is the genuinely unusual part.

You supply: One decision a dashboard changed. A dashboard nobody acted on is a depth probe you lose.

Before you press start

Read your own skills list aloud first. Naming a framework FLAGS it and generates a depth probe — this is the concrete reason nothing you cannot discuss belongs in that list.
One machine, and it also holds a client's material. Close everything else: the tracked-work software takes periodic screenshots and wants camera, mic and browser history.
Wired internet if you have it. A reconnect mid-answer is scored as the answer.
Silence beats a filler word. The transcript is what gets read, and "um" is in it — a two-second pause is not.

No profile exists until accepted. Five screening stages, under 3% accepted, 3–8 weeks, and a 1–3 week unpaid test project at the end. Rejection reportedly carries a six-month cooldown, so this is the last one to apply to, not the first.

Nothing to write here — that is the finding, not an omission.

ValueWhat it measuresSource

79% → 93% retrieval accuracy, measured research-console docs/revenue-research.md §4 · frozen

1,253 repositories evaluated research-console catalog/catalog.yaml

117 test modules, TDD research-console tests/

8 active projects governed research-console state/projects.yaml

~34k lines of Python research-console, wc -l over tracked .py files

$0 per-token API spend research-console — no API keys anywhere in the stack, by design · frozen

3,600 decisions in one evaluation auto-social-posting keyword pre-filter evaluation · frozen

8.8% word error rate, chosen voice ai-learning-path CLAUDE.md, edge-tts vi-VN-HoaiMyNeural bench · frozen

−80% average cost per install attested — VNG internal, not re-derivable · frozen

+30% monthly game revenue attested — VNG internal, not re-derivable · frozen

−80% manual workload, service and product teams attested — VNG internal, not re-derivable · frozen

A number with no source does not go in the record, because it does not belong on any of the surfaces either. “Frozen” means it carries no re-measure command — either it was measured inside a system we can no longer query, or deriving it is a manual job. Either way it is never reported as drift, because an alarm that always fires is not an alarm.

One record, 6 destinations

What the interview is

What it scores

Rehearse these 9

Before you press start