The CV of Adrien Vergne. M2 in NLP & AI at Sorbonne Université,
freelance engineering on the side. Python, machine learning, LLM pipelines, systems, hardware.
This is not a template: the whole OS (window manager, terminal, boot sequence) is one
hand-written vanilla JS file. View source, it all fits in there.
No AI assistant lives here anymore. The previous one (YIYI) was decommissioned in August 2026,
and is not missed.
Terminal tips: help · train · ollama run adrien · sudo hire adrien
about.txt
Adrien Vergne
AI & NLP Engineering · Python · Machine Learning · Applied Systems
Master's student at Sorbonne Université with training in software engineering, machine learning
and natural language processing, and a working practice built on real deliveries: data pipelines,
supervised classification, LLM inference and evaluation, ASR and OCR experimentation, API and
platform development, technical documentation, and embedded-oriented prototyping.
Working habits: models compared under identical conditions against gold data, safety red-teamed
instead of assumed, benchmarks reproducible, failures logged rather than hidden. Comfortable
anywhere between a LaTeX deliverable for an industrial client and a Linux driver, an IK solver
or a FreeCAD part.
Location
Saint-Maur-des-Fossés (Paris area), France
Status
M2 student · open to internships and missions
Focus
LLM pipelines · ML evaluation · systems engineering
Languages
French (native) · English (B2/C1) · Italian (B2)
~/projects
10 items · click a file to open it
p01_louvre.proj
AI & NLP · P-01
Louvre Catalogue LLM Pipeline
Master's thesis · Sorbonne Université
End-to-end normalization of 51,123 artifact records from the Louvre's public API: retrieval by
ARK identifier, LLM-based normalization of heterogeneous fields, geocoding and an interactive
map. Eleven models compared across 44 configurations (zero-shot, few-shot, NER pre-processing)
against a hand-annotated gold set of 300 records, with retry logic, controlled vocabularies
and separate error journaling so no failure passes silently.
▸ 0.961 macro-F1: a 31B model matching 675B mixture-of-experts
A read-only external auditor that independently recomputes the performance KPIs a Manufacturing
Execution System claims, from that system's own raw records, built to verify whether gains
promised in a subsidy dossier are real. Included a critical reading of a 130-page specification,
identification of indicators declared "unmeasurable", and the data-granularity questions that
decide whether KPIs are computable at all.
▸ 5 KPIs · measurement, not optimization
Systems analysis · KPI verification · LaTeX / TikZ · EN / KO
p03_smart-cafe.proj
Platform · P-03
Smart-Café Operating Platform
Multi-product platform · one account, one hub
A unified platform for café operations: in-store vision intelligence that adapts to existing
cameras and runs in real time on a CPU-only laptop, never exporting an image (privacy by
design); daily operations tooling (checklists, planning, stock, alerts); and a market
intelligence product built on public data. Three-plane architecture with contract membranes,
measured performance dossiers and reproducible cold-start demos.
▸ Real-time vision on CPU: aggregates only, no image leaves the device
Full record → train → eval pipeline with LeRobot and MuJoCo, run without a local GPU: recording
and evaluation on CPU, training on free cloud GPUs via code-first script kernels, checkpoints on
the Hugging Face Hub. It ends with a voice-driven vision-language-action model and a cockpit
app where a cloud VLM does the spatial reasoning while execution stays local, bounded by inverse
kinematics and safety guards.
Conversational AI for a vulnerable user, safety as the core spec
A multi-voice LLM pipeline (twelve specialist voices → architect → a final guardian pass) that
speaks its user's poetic universe without ever validating a harmful drift. The guardian is the
invariant: red-teamed against manipulation prompts, injected delusional corpora and escalating
model power, with the vital cases re-read by hand and tested for variance. Includes bounded
memory, RAG over personal documents, PBKDF2 auth and PWA packaging.
▸ 44–45 / 45 on the full red-team suite, 0 hard failures
TypeScript · Cloudflare Workers · Ollama · RAG
p06_care-network.proj
Platform · P-06
Care-Network Platform
Freelance mission · social-sector startup beta
A platform matching elderly, often isolated caregivers with younger helpers who visit regularly.
The bet: help becomes acceptable when it flows through transmission rather than assistance. Four user surfaces, geographic matching, a demonstration review site, and guard-rails
everywhere: static verification rules, HTTP leak checks, documented production pitfalls.
▸ 17-screen build with a static verification harness
SvelteKit · PostgreSQL · PostGIS · Node
p07_world-machine.proj
Simulation · P-07
Society Simulation Engine
Research prototype · a "world machine"
A simulation engine for an entire society (social, economic, political, ideological, cognitive)
that evolves on its own. Neither pure determinism (which drifts) nor pure LLM (which
hallucinates): every kernel, including the LLM, emits only typed pressures, resolved by
least-violation under a hard feasibility envelope, so plausibility is coerced by structure.
Exploratory exercise · explicitly not a prediction tool
A counterfactual simulation of a hypothetical 2027 presidential candidacy, calibrated on the
five previous French elections with leave-one-out validation, Latin-hypercube sensitivity
analysis and CMA-ES optimization. Includes an explicit audit of the model's own optimistic
biases; the corrected version honestly reports structural quasi-impossibility.
A Linux stack for Wi-Fi channel-state-information sensing on an Intel AX210: capture ingestion,
normalization into datasets, windowing, machine-learning baselines for presence and coarse
activity detection, and terminal/web dashboards, working at the interface of software, drivers
and radio hardware.
▸ Presence detection from radio signals alone
Python · Linux · Signal processing · scikit-learn
~/projects/misc
Multilingual ASR pipeline: audio preparation, Whisper transcription, model comparison and structured reporting on heterogeneous oral corpora.
Dementia detection from speech: binary classification on the DementiaBank Pitt corpus ("cookie theft" task).
Political classification on TEI-XML corpora: supervised TF-IDF pipelines with execution profiles, data audits and reproducible comparisons.
Zero-backend aid search engine: pre-computed Pagefind index served statically; search runs entirely in the browser, every result links to the official source.
Local PDF analysis pipeline: page-by-page extraction, graphic detection and LLM-assisted reporting on technical documents, fully offline.
French-learning game design: CEFR-calibrated engagement loop and LLM-tutor specification for a language-learning application.
RC parts compatibility engine: a PCPartPicker-style configurator for RC vehicles, with fitment rules (ESC ↔ motor ↔ LiPo ↔ chassis) as the core asset.
RC engineering: design, build and tuning of a 200+ km/h RC car; active-aero diffuser parts modeled in FreeCAD.
experience.log
2026freelance
Independent Engineering Missions
Technical missions for industrial and social-sector clients
Designed an independent MES performance-audit instrument for a Korean smart-factory subsidy
program: read-only recomputation of claimed KPIs from raw MES records, gap analysis of the
specification, and concise bilingual technical deliverables (LaTeX, English/Korean).
Built the beta of a care-network platform matching isolated elderly caregivers with younger
helpers: SvelteKit, PostgreSQL/PostGIS, four user surfaces, a static verification harness
and production-build checks.
Authored the product specification for a French-learning application: CEFR-calibrated
(A1→C2) engagement loop, LLM tutor design, and a full feature inventory consolidated from
two legacy codebases.
2024internship
NLP & Data Collection Internship
CERES · Sorbonne Université
Collected, structured and analyzed political and textual data from online sources.
Developed Python scripts for extraction, normalization and corpus-oriented analysis.
Worked on categorization, data representation and OCR over documentary corpora.
2023university
Team Project Coordination
Sorbonne Université
Coordinated a 3-person team on a design and implementation project: task allocation,
progress tracking and technical scoping.
skills.sys
Languages
Python · TypeScript / JavaScript · R · C++ · Java · SQL · LaTeX
Master's student at Sorbonne Université (M2, NLP & AI applications) with training in software
engineering, machine learning and natural language processing. Real deliveries across data
pipelines, supervised classification, LLM inference and evaluation, ASR/OCR, API and platform
development, and embedded-oriented prototyping. Every claim measured: gold-data evaluation,
reproducible benchmarks, red-teamed safety, logged failures.
Experience
2026 · Independent engineering missions (freelance). MES performance-audit
instrument for a Korean smart-factory subsidy program (read-only KPI recomputation, bilingual
EN/KO LaTeX deliverables); beta of a care-network platform (SvelteKit, PostgreSQL/PostGIS, four
surfaces, verification harness); product specification for a CEFR-calibrated French-learning
application.
2024 · NLP & data collection internship, CERES (Sorbonne Université).
Collection, structuring and analysis of political and textual data; Python scripts for
extraction, normalization and corpus analysis; categorization and OCR on documentary corpora.
2023 · Team project coordination, Sorbonne Université. Coordinated a 3-person
design-and-implementation project.
Selected projects
Louvre catalogue LLM pipeline (Master's thesis): 51,123 records normalized
via the Louvre public API; 11 LLMs × 44 configurations vs a hand-annotated gold set;
0.961 macro-F1; geocoded interactive map.
MES performance audit: independent read-only recomputation of 5 claimed
KPIs from raw MES records for an industrial subsidy dossier.
Society simulation engine: multi-kernel "world machine": typed pressures
resolved by least-violation under a feasibility envelope.
Electoral counterfactual model: calibrated on 5 French elections,
leave-one-out MAE 4.43 pts, explicit bias audit.
Wi-Fi sensing / CSI stack: presence detection from radio signals on Linux
/ Intel AX210.
Also: multilingual ASR pipeline · dementia detection from speech (DementiaBank Pitt) ·
TEI-XML political classification · zero-backend aid search engine · local PDF analysis
pipeline · French-learning game design · RC compatibility engine · 200+ km/h RC build with
FreeCAD active-aero parts.