Full-stack developer, Java / React and applied AI. Final-year engineering student at CY Tech (graduating 2027).
Looking for a six-month end-of-studies internship from 4 May 2027, in or around Paris.
As the sole developer, orchestrating AI agents, I designed and shipped the production-management system of a leather factory (MES / IIoT). It collects machine readings, makes them reliable, computes output and piece-rate pay, and presents it all in eight web and Android apps.
Simulation: each line is a window of readings from a spray machine, and the newest is written at the bottom. A suspicious reading is never silently fixed: it is flagged in red, with the anomaly code. Tap or click the field to inject one.
The journey of one reading, from machine to screen, in six steps:
Each machine has an industrial touch screen (HMI) wired to its PLC. I defined the format of the messages these screens send (MQTT protocol) and each machine’s identity; the factory’s engineers configure the screens to that spec.
device/{clientId}/spray device/{clientId}/energy-points
The broker (Mosquitto) receives the machines’ messages and passes them to the back end. Guaranteed delivery (QoS 1) and persistent sessions: if the back end stops, the broker holds up to 20,000 messages, about 80 minutes of downtime absorbed without loss.
cleanSession=false: the session resumes where it left off.The back end acknowledges a message to the broker only once it is stored in the database. If the write fails, the message is not acknowledged and the broker sends it again. A message received twice is counted once.
client|seq|deviceTime: a resend is recognised as a duplicate; the same key with different content is rejected.A suspicious reading is not corrected: it gets flags that say why. For example: an impossible or physically out-of-range value, special code, missing or duplicate message, clock off by more than 10 minutes. The raw value is kept, but left out of the statistics.
PostgreSQL 16 with TimescaleDB, a time-series extension. Hourly and daily totals are recomputed only where a reading changed, and screens update in real time (WebSocket). A query that is too heavy fails rather than return an incomplete total.
One API contract (openapi.json) generates the Kotlin and TypeScript client code of every app: if the API changes, the Android apps stop compiling until they catch up. A business rule: machine readings never create a production report or a wage. The machine measures; only a person declares what gets paid.
The context: the factory sprays finishes onto hides and pays its workers by the piece. It wants to know, machine by machine and shift by shift, what was produced, consumed and paid. The difficulty: the machines’ industrial screens drop messages, sometimes impersonate one another and switch units, while payroll tolerates no approximation.
Figures as of 12 October 2026.
WhyA message is lost only if the database refused it, and then the broker sends it again.
What it cost meEvery resend creates a duplicate to recognise: each message needs a uniqueness key, and a clear rule for when the same key comes back with different content.
WhyA rewritten value is destroyed evidence. When a cleaning rule changes, the whole history must be recomputable.
What it cost meStatistics are computed on read or rebuilt. Every algorithm change means recomputing history, after a rehearsal on a copy of production.
WhyA missing counter is not a zero. Payroll and production totals cannot rest on an estimate.
What it cost meSome totals stay “partial”, and the interface has to show it: a less flattering dashboard, but a correct one.
WhyEight apps, each moving at its own pace, must understand the server in exactly the same way.
What it cost meEvery field change means regenerating code and rebuilding the Android modules. On the TypeScript side, the tool (springdoc) does not mark required fields: the shape of the data is checked, not its presence. That annotation work is still to do.
WhyAgents produce faster than the full test suite runs: waiting for it on every merge would have stalled the work.
What it cost meThe main branch can stay red for a while; we fix forward. Only a version that passes the whole suite, on a dedicated test machine, can go to production.
An industrial system is judged on its real data. The rule: recompute history, never touch a raw reading.
spray-interval-v4), leaving the raw readings untouched, then history recomputed. Rehearsed first on a copy of production: 1,363 hourly rows and 75 production runs recomputed in about 4 minutes.usedWeight) held not a running total but a difference: W0 − totalWeight.A machine sends more readings while it works than while it is idle. An average taken reading by reading therefore gives too much weight to busy periods, and it shifts as soon as fewer readings are stored. Weighted by how long each reading lasts, the average stays correct. Drag the slider to see it.
Reading by reading
Time-weighted
True value
| Storage | Reading by reading | Time-weighted |
|---|---|---|
| 1 reading / 2 s | 783.9 +46.5% | 535.1 0.0% |
| 1 reading / 10 s | 667.9 +24.8% | 535.1 0.0% |
| 1 reading / 60 s | 518.0 −3.2% | 535.1 0.0% |
Three systems, one rule: the server checks every number, dish and correction before it reaches the user.
The owner asks a question; the agent answers with charts in which the model writes no number.
I don’t use an agent framework: I write the loop myself, and I limit what the model is allowed to do.
The factory owner asks a question in plain language. The agent queries the database read-only and answers with charts whose every number comes from the database, not from the model. The same agent serves as the product’s built-in help, that is, its customer support.
SELECT machine, SUM(kwh) AS kwh FROM v_energy_day WHERE day >= current_date - 7 GROUP BY machine ORDER BY kwh DESC12 rows, complete → evidence-1{ evidenceId: "evidence-1", type: "BAR" }the server reads the values from evidence-1; factory ✓ unit ✓It answers customers from the restaurant’s real data; dishes and allergens are checked by code.
A restaurant’s customers write on WhatsApp or on the website: opening hours, menu, allergens, bookings. The assistant answers in their language from the restaurant’s real data, and hands over to the staff when needed.
Hi, are you open on Sunday lunchtime?
Hello! I’m the restaurant’s AI assistant. Yes, Sunday from 12:00 to 14:30. Shall I book you a table?
BookSee the menu
rateinputquotacachebudgetmodel
opening time checked against the restaurant’s data
Do you have a gluten-free dish?
Yes: the porcini risotto and the grilled salmon are marked gluten-free by the restaurant.
See the menu
rateinputquotacachebudgetmodel
dishes on sale only, per the restaurant’s own labels
Am I talking to a real person?
No, I’m an AI. I’ll let the team know: someone will reply during opening hours.
Call the restaurant
rateinputquotacachebudgetmodel
mandatory AI notice, handed over to a human
The AI corrects a sentence within seconds; its quality was measured against thresholds set in advance.
A vocabulary app for Chinese-speaking students in France. The learner writes a sentence with the word of the day; the AI corrects it within seconds and explains the mistakes in Chinese. No grade and no opinions on style: only objective errors.
Word of the day emprunter (to borrow)
Hier, j’ai emprunté deux livre à la bibliothèque.
Hier, j’ai emprunté deux livres à la bibliothèque.
名词复数要加 -s：deux livres。 (“A plural noun takes an -s.”)
Use of the word of the day: correct
| Criterion | Threshold | Without cards | With cards |
|---|---|---|---|
| Word use judged correctly | ≥ 90% | 94.6% | 95.2% |
| Wrong use accepted as correct | ≤ 5% | 3.6% | 1.8% |
| Unusable answers | ≤ 1% | 0% | 0% |
| Response time (95% of cases) | ≤ 3.2 s | 3.16 s | 2.12 s |
| Prompt served from cache | – | 0% | 76% |
| Cost of the run | – | $0.024 | $0.013 |
First run with cards: 92.8%. The model mistook scenario details for a change of meaning; I fixed two sentences of the instructions, ran it again, and kept all three runs. The test sentences are synthetic: this is not teacher grading.
AI did not mainly make me code faster. It freed my time for what cannot be delegated, deciding:
About 3,900 commits and 49 releases in five months as the sole developer, with up to some twenty agents in parallel, each in its own copy of the repository. With agents, volume proves nothing: what counts is what tests and production confirm.
How a change travels from an agent to production:
main branchthe full test suite starts on a dedicated test machinemain-verifiedonly advances to a commit where every test passesThe instruction file read at every session holds only the costly misunderstandings and pointers; everything else loads on demand.
If a rule can be checked automatically, it becomes a hook, an architecture test or a consistency check. A rule you have to keep repeating to agents is in the wrong place.
A good test judges the code from the outside instead of mirroring it: that is what stops an agent from approving its own work.
Rules are written once, then generated for each agent tool; a checker refuses any drift. The same principle as the API contract, applied to rules.
Do my reusable instructions (“skills”) really change the agent’s behaviour? An A/B test with a protocol fixed before measuring, in a sandbox with a real Git remote. Verdicts come from the state of the repository, not from the agent’s own account; placebo scenarios and blind review.
With the skilltests the error cases, waits for approval
Withoutdeclares itself “done” and commits directly
Verdictslight advantage
With the skillidentical result
Withoutidentical, byte for byte
Verdictskill redundant, retired
With the skillread the rule, then pushed to the main branch anyway
Withoutopened a branch
Verdictrule turned into an automatic block (hook)
Result: +26% tokens per run, for an inconsistent gain with an already strong model. Two skills retired; what must hold moved into a blocking hook. A pilot with one run per variant: a signal, not a statistic.
May 2026 →
Full-stack developer, sole engineer on the whole software, independent contract (freelance)
Production-management and connected-machine software for the leather industry, in production at a manufacturer. From the machine protocol to the apps.
2026 →
Full-stack developer, part-time permanent contract, after a five-month internship
Ordering and loyalty platform for restaurants. I build the WhatsApp AI assistant, part of the back office (React, TypeScript) and field deployment: servers installed automatically in restaurants, Android tablets in kiosk mode, and a self-hosted operations centre to monitor and update the fleet remotely.
2026 →
Personal project, live
French vocabulary app with AI sentence correction; 201 lessons online.
2024 → 2027
Engineering degree in computer science, embedded systems track
Taught entirely in French.
Sept 2021 → Apr 2024
Front-end and Android developer
Web interfaces for 3D scanning equipment (Vue, JavaScript); a face-recognition payment terminal (Android, Java).
2017 → 2021
Bachelor’s in electronic and information engineering, Sino-French programme
A six-month end-of-studies internship from 4 May 2027, in or around Paris, in full-stack development, and I am open to a permanent position after the internship. Ideally in a team close to its customers, that uses AI to design and not only to execute. I like to own a scope early: that is how I have learned the most.