Théo CHARLET (RDTvlokip)

About

"I am dyslexic. And that's my superpower. It allows me to approach complex problems with a unique and creative perspective."

Tokenizer combining BPE with weighted multi-layer Transformer attention (v4) for semantically coherent merge decisions. 3.85x compression, 0.03ms decoding, zero hard OOV on multilingual text (French, Korean, math symbols, emojis).

RDTextract

HTML→Markdown extractor designed for LLM training corpora. Eliminates all noise (navigation, ads, repeated widgets) with an integrated quality scoring system and a JSON-LD fallback for SPA pages. Score of 99.3/100 on the Tranco top-1,000 .fr benchmark — zero artifacts on all 4 measured metrics. Multi-language stub detection (FR/EN/ES/DE/IT). One dependency: beautifulsoup4.

New security paradigm — Destructive Semantic Encoding (DSE) — distinct from classical cryptography, hashing, and compression. 7 hierarchical levels of semantic scrambling culminating in a random alphabet mutation unique to each compression (26! permutations). >10^35 years brute-force resistance, tested on 40,000+ French phrases.

Independent search engine focused on privacy and transparency. From-scratch crawler (RDTvlokipBot), adaptive BM25 + PageRank algorithm (24 scoring signals), 5,021,846+ indexed pages, hosted in Germany — zero ad trackers.

RDTSearch CLI

Official terminal client for RDTvlokip Search, my independent, privacy-first French search engine (~5,021,846 indexed pages, in-house crawler). Direct search in <200ms, smart autocomplete, tracker and adult-content detection, live engine stats, full SEO/backlink analysis (link graph, domain duels, CSV/JSON export) and a full interactive TUI. Single-file, zero-dependency Python 3.8+ tool (stdlib only), bilingual French/English, Windows-first (Linux/macOS supported), released under the AGPLv3 license.

Family of three 15M-parameter French LLMs trained from scratch (no pretrained weights, no dependency on 🤗 Transformers) on a single GTX 1080 Ti, exploring looped (recurrent-depth) transformers and per-token adaptive computation. Cadence applies the 8 transformer blocks 4 times (effective depth 32) with no early exit; Focal and Nomade add per-token early exit, using an absolute entropy threshold and a size-invariant percentile threshold respectively. Result: -7% validation perplexity for Cadence (28.9 vs 31.2 baseline), but no single model wins on every axis.

Experimental testbed for whether a neural network can learn to produce language from a scalar reward alone, with random initial weights, no pretraining, and no input/output pairs. Search spaces are small enough to enumerate exhaustively, so valid-mass, effective modes, the Gibbs optimum, and conditional structure are computed exactly rather than estimated. Test 1 (copying a fixed sentence) shows success comes from reward decomposability, not search itself. Test 2 (a hand-written grammar judged by a deterministic parser) reaches 99.9% grammaticality with zero learned rules — but both "successes" turn out to be empty on inspection. Test 3 (a two-agent referential game) concludes that only the network's parametrization produces compositionality, never selected by training itself. v0.5.0, five rounds of external review by Dipankar Sarkar.

An 11.05M-parameter residual network (128 channels × 8 blocks, recurrent depth ×2, gated memory, Squeeze-and-Excitation) that plays chess in one forward pass — no search, no MCTS, no opening book. This is RDTChess2, the model currently live on the Lichess bot. Trained from scratch on a single GTX 1080 Ti by behavioral cloning on ≈10.99M Lichess games (both players ≥1800 Elo), seeing 801,295,726 positions. 61.5% first-move accuracy on 23,845 Lichess puzzles, but the network still executes tactics rather than initiating them: the quiet entry move remains the weak point, at 33.3% in hard positions versus ~90% on the final step of a forcing sequence.

A native technical markup language and VS Code extension, built entirely from scratch with zero runtime dependencies. It replaces the fragmented Markdown ecosystem by bringing LaTeX mathematics, interactive charts, structured diagrams, tabs, metric dashboards, conversations and modular files into a single specification. Everything is generated as SVG and DOM, with no innerHTML or eval, so it works under the strictest Content Security Policies (CSP).

Collaboration with LMC. Development of a GPT-2 tokenizer optimized for large-scale natural language processing.

Uptime monitor for Discord bots and your websites/APIs. Instant downtime detection via Presence intent, HTTPS monitoring with URL security verdicts (anti-phishing, TLS certificate, domain age), customizable alert embeds, instability detection and full web dashboard.

AI & NLP Research

Systems & Networks (TSSR)

Development & Tools

Languages

AI & Research

Infrastructure

Development

VTT Editor Pro

Advanced WebVTT subtitle editor. Timeline visualization, auto-correction and multi-platform compatible export.

AG-BPE

Try AG-BPE directly.

RDTvlokip Search

The only search engine that shows you why each result appears, lets you control the ranking, and proves it doesn't spy on you.

RDTextract

HTML→Markdown extractor for LLM corpora. Zero artifacts, built-in quality scoring, one dependency. Score 99.3/100 on Tranco top-1,000 .fr.

WhisperWatch

Monitor your Discord bots and websites 24/7. Instant alerts, URL security verdicts, customizable embeds and full dashboard. Premium free for early adopters.

What I'm building

• From-scratch crawler, no Bing dependency
• 24 ranking signals — Adaptive BM25 + PageRank
• 5,021,846+ pages indexed solo, on Hetzner VPS + GTX 1080 Ti

Recently

• ✓TSSR Graduate — Nepsod, Bel Air School
• ✓LISP-7 published on Zenodo
• ✓AG-BPE — 1K+ downloads