Skip to content
ELM Labs

Work

Selected work from an applied AI lab — detection, AI agents, prediction, and autonomous pipelines, plus the research behind them.

AI & Security

Detection systems and AI agents — the work closest to what we focus on today.

Customer Anomaly Detection System

Large manufacturing company

Production data isn't Gaussian. This model fits a custom probability distribution to the real data shape, then flags anomalies against it. 56,234 records, 6 interactive tabs.

Fits a custom probability distribution to real production data, then flags what does not belong — the core of detection and monitoring.

1Raw Signal Ingestion
seasonality
raw signal
heavy tail
2Distribution Analysis
Gaussian rejected
actual shape
≠ 0
skewness
> 3
kurtosis
< 0.01
p-value
3Custom Distribution Fit
μheavy tail
shape αscale βloc μ
4Rolling Z-Score Engine
+σ−σ
sliding window scores each point
outlier
5Anomaly Classification
True anomaly
12%
Statistical noise
68%
Edge case
20%
Actionable alerts only
56,234 records80 months6-tab interactive dashboard
PythonBI DashboardData AnalysisPostgreSQLStatistics

OnePilot

Developer

Your server, in your pocket. SSH into any machine, deploy AI agents, manage files, git, and cron jobs — all from your iPhone.

Deploy and operate AI agents on real infrastructure — securely, from anywhere.

23+ LLM providers3 messaging channelsFull mobile IDE
SwiftUISSHVT100WebSocketSupabase

ML & Prediction

Models that turn noisy data into probabilistic, decision-ready signals.

Trading Regime Models

Individual quant trader

Time-series data shifts between hidden regimes. This system detects two layers — the macro state and the daily signal conditioned on it. 67.3% next-regime accuracy.

Two-layer clustering surfaces hidden market regimes from raw price data.

1Price Data Ingestion
bullish
bearish
OHLCV
2Feature Engineering
f(x)ReturnsVolatilityVolume ΔMomentum
raw price → feature vector
3Macro Regime Detection
volatility →return →
Trending
Mean-reverting
Volatile
4Daily Regime Classification
20 days
bullish
neutral
bearish
5Adaptive Position Sizing
Trending ↑
1.0x
Mean-revert
0.5x
Volatile ↕
0.25x
Trending ↓
0.8x
Size adapts to detected regime
67.3% next-regime accuracyXGBoost + KMeansLive signal
PythonMachine LearningClusteringPostgreSQL

WC 2026 Prediction Model

Betting professional

Three methodologies where one isn't enough. Statistical model + ML feature layer + 50,000 Monte Carlo tournament simulations.

Statistical model + ML feature layer + 50,000 Monte Carlo simulations for probabilistic forecasts.

1Historical Match Data
FRA
2–1
GER
BRA
1–1
ARG
ESP
3–0
ENG
ITA
0–2
POR
ARG
3–3
FRA
GER
1–2
ESP
10,000+ historical matchesFIFA 2002–2026
2Dixon-Coles Model
01234Standard Poisson
correction01234Dixon-Coles
τ corrects 0-0 and 1-1 scores
3Attack/Defense Ratings
attackdefense
ESP
ARG
FRA
BRA
ENG
α (attack) and β (defense) parameters per team
4Monte Carlo Simulation
R16QFSFF3rdW
0K simulations
each path = one tournament
5Win Probabilities
1ESP
18.7%
2ARG
16.4%
3FRA
14.3%
4BRA
12.1%
5ENG
9.8%
Updated after each matchday
5,030 matches50,000 Monte Carlo runsDixon-Coles + XGBoost
PythonStatistical ModelingMachine LearningSimulation

Data & Automation

Pipelines that source, transform, and act on data without a human in the loop.

Prospector

B2B sales team

Automated pipeline: discover businesses, normalize messy real-world data, deduplicate, score, and reach out. End to end, no manual step.

End-to-end pipeline: discover, normalize, deduplicate, score, and reach out — no manual step.

Pipeline Architecture

1
prosp_source
prosp_source_sub
prosp_source_detail
2
prosp_normalize
prosp_normalize_sub
prosp_normalize_detail
in:06 12 34 56 78
out:+33 6 12 34 56 78
3
prosp_deduplicate
prosp_deduplicate_sub
prosp_deduplicate_detail
4
prosp_score
prosp_score_sub
prosp_score_detail
prosp_score_high
85%
prosp_score_medium
55%
prosp_score_filtered
20%
5
prosp_deliver
prosp_deliver_sub
prosp_deliver_detail
ACTIVE1 SMS/min · budget tracked
9 cities × 10 categories90 search combosGoogle Places + Pages Jaunes
PythonAPIsWeb ScrapingSMSPostgreSQL

Eurostat Market Intelligence

Industrial distributor

Raw government data is fragmented across APIs, product codes, and formats. This pipeline stitches it into a market picture — supply, demand, trade flows — automatically, monthly.

Stitches fragmented government data into a monthly market picture, automatically.

Comext + STS + yfinanceMonthly automatedIndustry-grade B2B
PythonPublic APIsData ProcessingPostgreSQL

Research

The recurring theses behind the projects above — how we think about signal, autonomy, and the bridge between systems and people.

Signal

Finding structure in noisy data through layered statistical methods.

Architecture

Raw data→Statistical layer→Signal

Autonomy

Systems that source, transform, and act on data without human intervention.

Pipeline

Source→Transform→Validate→Deliver

Bridge

Connecting complex systems to human understanding across platforms and devices.

Protocol

Complex system→Protocol layer→Human interface