Not a demo.
ProspectIQ enriches B2B sales prospects with AI. Colleagues at FPT use it in their real sales workflow — a tool people actually depend on.
Paris, France — 48.8566°NAs featured on Vietnam's national TV
Product Engineer — I design and build AI products end to end.
I work at the collision of design and engineering — where most people pick a side.
01 — Case studyIn production · used inside FPT Software France
ProspectIQ
An AI-powered B2B prospecting platform, actively used inside FPT Software France — rebuilt to survive the death of its core data vendor.
A three-front fire.
The scraping-based data source died — Proxycurl shut down after the LinkedIn lawsuit. Some enrichment data was silently wrong. And the original approach didn't scale. A live, depended-on tool was about to break.
Cross-model orchestration.
The pipeline runs Claude + Gemini — and API rate limits kept breaking the hand-off between them. Licensed-data cost and matching quality were a war of their own.
Built for the failure case.
Multi-provider with fallback — no single data vendor can take the system down. A queue + retry layer for the model calls — no API ceiling can break the pipeline.
Then the vendor actually died.
When Proxycurl shut down for good, the tool colleagues depend on kept running. The thesis paid off in reality — and the system is still actively hardened today.
“Build for the failure case first — vendors will die on you.”
~95%
cheaper than per-seat prospecting tools
Replaces per-seat licences (Sales Navigator, Apollo) with API cost per dossier.
02 — Case studyLive · no sign-up · target7.io
Target7
An essay grader whose scores land as close to a human examiner's as two examiners land to each other.
On a held-out set of examiner-marked essays, agreement reached |Δ| ≈ 0.45, approaching the inter-rater agreement between two human examiners.
Small n (≈29), corpus skewed mid-band — extreme-band anchors are next. Knowing a metric's limits is part of the work.
Human IELTS grading is slow and expensive. Target7 grades instantly, with real feedback — not just a number. In production at ByPath Edu.
Not “ChatGPT with a prompt”
- Deterministic
- Temperature 0 — the same essay gets the same score, every time. Rare and hard for LLM grading.
- Calibrated
- Examiner re-anchoring + isotonic regression to correct individual rater bias.
- Architecture
- DeepSeek + Gemini scoring, two-pass ensemble, ~10k-token system prompt of atomic grading rules, handwritten-OCR intake.
~99%
cheaper per grading than a human tutor
Interlude — Thăng LongVN · FR · EN
The ascending dragon.
Hà Nội's old name is Thăng Long — “ascending dragon.” It's my name too. Paris ⇄ Hà Nội: the route behind Skyhop, and behind everything else I build.
48.8566°N 2.3522°E ⇄ 21.0285°N 105.8542°E
Take the flight03 — More builds
Range, proven
PDFForge
0126 PDF tools where your files never touch a server — privacy isn't a setting, it's the architecture.
Solo-built, fully client-side. A free alternative to €84–240/yr tools.
Privacy by architectureOpen the tools
Skyhop
02A flight finder for Paris ↔ Vietnam that tags every route by airspace safety — filter out conflict-zone airspace.
Defensive parsers, tiered scrapers, a mock fallback so the UI never goes empty. The failure-first thesis — proven a second time.
The resilience pattern, againPrivate demo
CANAL+
03UX audit — heatmaps, navigation analysis — delivered as a full design system: tokens, components, docs.
Real audit method in, shipped deliverable out, for a brand every French recruiter knows.
Design at client grade
Orange Brain
04A learning PWA built as a complete design system — Nothing-OS aesthetic, end to end.
Its engine adapts the curriculum to what you're actually building.
Design-systems range
Plus a handful of smaller builds — a real-estate aggregator, a watch-together streaming app, and other tools I've shipped for people around me.
04 — About
The collision
Most people pick a side — design or engineering. I never could. I build at the seam between them, because that's where the interesting problems live and where most teams have a gap. It means I can take something from a user interview to a deployed, AI-powered product without a hand-off losing the thread. ProspectIQ runs inside FPT because of that. Target7 grades essays like a human examiner because of that. I learn whatever the problem needs — calibration math, quantization, a new vendor's API — to ship the thing properly.
Before the AI work, I ran cross-regional UX research at FPT — 15 participants, interviews across Europe and Asia — so the “design” in design-engineering isn't decorative. And my first “product” was growth: four years running digital for my family's restaurant, Kaizen, in Germany — from 3.8 to 4.5 stars and 1,400+ reviews. That's where I learned that shipping means moving a real metric, not just launching something.
- 01Trilingual — VN · FR · EN
- 02Master's UX Design — Efrei Paris / Panthéon-Assas
- 03BSc Computer Science, Honours — Université Paris Cité
- 04Alternance — FPT Software France
- 05Mentors junior React devs
- 06Featured on Vietnam's national TV