DATA / AI / SYSTEMS
Achraf Ez
Data & AI Consultant · Agentic Entrepreneur · Founder
I like solving problems. The bigger the problem, the stronger my determination.
Paris, France
- Production ML
- Time series
- Agentic systems
- Open source
Who am I?
I build data and AI systems from a foundation in production machine learning, time series, and engineering.
Selected work
Pipeline Video
Turns a company's website into a finished marketing video: crawl, script, human approval, recorded voiceover, motion graphics, deterministic render.
- A multi-stage LLM pipeline in production, with every model call site behind one provider-agnostic protocol.
- A deterministic floor that guarantees delivery when the model pass fails.
- A measurement harness that cut end-to-end wall clock 49% across seventeen instrumented runs.
- Deployed into production and currently serving users.
LumiereAI
Autonomous agents for enterprise operations.
auditMalin
AI-assisted Qualiopi pre-audits across the RNQ's 32 indicators, with cited verdicts and prioritized action plans.
- A multi-stage agentic RAG pipeline in production, with a provider-agnostic model layer behind one environment switch.
- A deterministic floor that downgrades ungrounded model output to not verified rather than guessing.
- Criterion-level batching that cut 32 sequential LLM calls to about 7 concurrent ones, a 4.5x reduction in call count.
- Deployed into production and currently serving users.
FidelityAI
Explainable multi-agent insurance claim triage across documents, photos, voice statements, and policy evidence.
- A multi-layer image-forensics engine combining seven classical and deep-learning signals with an LLM judge under a strict structured-output schema.
- Benchmarked at 80% accuracy with 100% precision on hard negatives.
- Recall improved by about 80% relative through threshold recalibration driven by error analysis.
- Two prizes at the Pioneers hackathon at Station F, including best overall.
Open source
Contributed to improve time-series diagnostics around missing data.
Contributed to improve Fourier smoothing for real-valued signals.
Contributed to improve unequal-length time-series workflows.
Contributed to improve wording in model-explanation outputs.
Contributed to improve robust categorical encoding workflows.
Contributed to improve datetime feature engineering.
Contributed to improve seasonal period detection.
Contributed to improve reliable package installation.
Contributed to improve calendar features for time-series forecasting.
Contributed to improve edge-case padding behavior.
Contributed to improve frequency-domain anomaly detection.
Contributed to improve spectral anomaly-detection workflows.
Contributed to improve calendar-aware forecasting features.
Contributed to improve reliable model-selection workflows.
Writing
I Stopped Prompting Claude Code and Started Engineering the Loop
A green test suite missed an integration seam; a frozen mutation experiment shows how repository rules, executable checks, scoped review, and replay state make agentic work more inspectable.
Aeon Periodograms for Unequal-Length Time Series
How Aeon's PeriodogramTransformer pads unequal-length collections, why its list and 3-D branches need parity, and what one case study teaches.
Fourier Smoothing in Aeon: Symmetry, Cutoffs, and Better Tests
A bounded Aeon DiscreteFourierApproximation case study shows how conjugate symmetry, passband gain, and cutoff semantics expose concrete Fourier failure modes.
Your ACF Can Be Right While Its Confidence Interval Is Wrong
A missing-data bug in statsmodels left ACF estimates unchanged while confidence intervals and Ljung-Box statistics used the wrong sample size.
One Prompt, Thirty-Two Calls, or Seven
Batching an LLM checklist is a blast-radius decision before it is a performance one: a seeded simulation of three call shapes, and why a call count is arithmetic while a speed-up is a measurement.
A Detector With Perfect Precision and Zero Recall
A document check cleared every genuine file and every forgery in a labelled set; a seeded synthetic run shows why specificity on genuine inputs is free and recall is the first number to ask for.
The Constant Nobody Could Re-Measure
A magic number survives because there is nothing to re-measure it with. The calibration data is often already in a log you throw away, and the same data sets the noise floor that makes the result readable.
The Forecast Metric That Hides the Failure
A volume weighted error of 7.6 percent and a net bias of 0.2 percent can describe a forecast that loses to a seasonal naive on 77 of 116 slow moving series. A seeded panel shows what to measure instead.
From Brief to Rendered Animation: Making an AI Media Pipeline Inspectable
A failed scene edit shows why reliable AI video systems need approved scripts, timed artifacts, deterministic fallbacks, and media read-back.
Calendar Features for TabPFN Time Series: Geometry, Encodings, and an Open Edge Case
TabPFN-TS turns forecasting into a table-plus-features problem. A 24-hour edge case shows why geometry, checkpoint compatibility, and release evidence are separate gates.
How Far Should You Tune LightGBM? A Budget for the Last Fraction of a Point
LightGBM tuning can win, but only a deployment-faithful baseline and nested validation can show whether the last fraction of a point pays for its search.
Are Tabular Foundation Models Worth the Hype? My Bet Is Yes - but Not for the Reason You Think
Tabular foundation models are a real shift, but the durable bet is about information sets, task breadth, and synthetic priors - not easy benchmark wins.
Insurance AI Needs Evidence Graphs, Not Just Fraud Scores
A synthetic insurance claim shows how hashes, typed facts, contradictions, abstention, and human review make AI recommendations auditable.
From PDFs to Citable Verdicts: Engineering an Evidence-Grounded Qualiopi Pre-Audit
A practical design for separating relevance from proof, preserving document lineage, abstaining safely, and evaluating a citable Qualiopi pre-audit.