"Connect the data backwards, invent the future forwards, and stay curious."

Data Scientist with 3+ years of professional experience in ML, NLP, and data engineering, now focused on LLM-based agent systems — making them behave predictably with typed contracts, explicit state, and evaluations that measure a claim instead of assuming it.

Data science & ML (professional)

BI & reporting

Backend & AI agents (personal projects)

Tooling

• TrainFitter, Twistify, and AuraPulse — three portfolio projects moving phase by phase rather than shipped-and-done, each covering a deliberately different slice of the agent-engineering roadmap: a multi-agent pipeline with a human-approval gate (TrainFitter), a measured-not-promised safety guarantee with a calibrated eval harness (Twistify), and conditional routing / when a graph orchestrator earns its complexity over a plain sequential pipeline (AuraPulse).
• Deepening a specific set of AI-engineering topics, in this order of priority right now:
  • Agent architecture — typed tool contracts, explicit state, bounded planning, permissions, retries, human approval.
  • Context engineering — loading only the context that's relevant when it's needed, instead of stuffing the whole prompt.
  • Production RAG — ingestion, chunking, embeddings, hybrid search, permission-aware filtering, reranking, citations, evaluation.
  • Evals & observability — purpose-built test datasets, traces, success rates, latency, cost per task, tool errors, regressions.
  • Efficient fine-tuning — SFT + LoRA/QLoRA, DPO/GRPO, only once there's real data and a clear metric — not as a first resort.
  • Inference infrastructure — vLLM, batching, semantic caching, small models for subtasks, routing between models.
  • Security — prompt injection, tool isolation, secrets handling, per-user authorization, audit trails.

TrainFitter A multi-agent system that drafts workout and nutrition plans for a personal trainer's clients, following the trainer's own documented method instead of generic advice — now with a live demo, Gmail/Notion integrations, a client-facing portal, and a trainer-facing client roster, not just a pipeline. Problem it solves: the bottleneck in online coaching isn't coaching — it's the hours spent writing a routine and a diet from scratch per client, then tracking whether the client actually follows it. Stack: Python 3.12, a deterministic rules engine as the free default path (with an optional Anthropic Claude layer behind the same schema), explicit-state orchestration across routine/diet/validator agents, a Streamlit review panel and client portal, Gmail/Notion connectors, a GitHub Actions cron trigger, pytest, CI. Notable design choice: nothing is ever sent to a client automatically (one narrow, disclosed exception for the portal's own magic link) — every plan is a draft, and clinical or injury cases are auto-flagged for human review by a validator that's deliberately never the LLM path. Recent improvement: a client roster with per-client weight/adherence trend charts, surfaced by a real production crash (a missing chart-library dependency) that live-testing caught and the test suite hadn't. Try it: trainfitter.streamlit.app — no install, no login, no API key.

Twistify A spoiler-free movie catalogue where the spoiler partition is enforced server-side, paired with an evaluation harness that measures whether that promise actually holds — including the uncomfortable case where the cheap judge fails. Problem it solves: "spoiler-safe" is usually a UI trick (CSS hiding a div); here, post-viewing content simply isn't sent to the client until it declares seen=true. Stack: Python 3.12, FastAPI, Pydantic v2, vanilla HTML/CSS/JS on the frontend, two interchangeable baseline generators (Anthropic paid / Groq free tier), a custom evals harness (leakage rate, grounded-fact rate, richness) calibrated against a 7,657-review external human dataset, pytest, CI. Notable design choice: the harness reports its own judge's weaknesses (a measured recall = 0.089 on the best judge so far) instead of hiding them, and blocks the next milestone (retrieval) until the judge clears a trust bar. Recent improvement: a research-assist tool that drafts new catalogue entries from real Wikipedia/TMDB retrieval (never LLM memory), with a code-level safety net that strips any fabricated citation — confirmed working end-to-end live, best-of-3 candidate generation included. Try it: twistify.onrender.com — 8/20 titles fully researched with cited sources, the rest browsable via TMDB.

AuraPulse An agent that reads a restaurant's public reviews and detects recurring operational inconsistencies (food praised while wait time is consistently criticized, for example) instead of just reporting aggregate sentiment. Problem it solves: sentiment dashboards tell a business "you're at 4 stars"; this turns the same reviews into a specific, actionable signal about what is quietly degrading. Stack: Python 3.11+, Pydantic v2 schema-first classification, a local LLM served by Ollama (zero paid API calls anywhere in the project), pandas for dataset handling, pytest/ruff/mypy in CI. Notable design choice: ground truth is never LLM-generated — a deterministic fake-review generator validates the pipeline first, and the free Yelp star rating backs sentiment evals before a single review gets hand-labeled for aspect extraction. Status: Hito 0 (classification → aggregation → reporting) done and evaluated end-to-end; Hito 1's first slice has since shipped too — LLM-free routing, draft-reply generation, and deterministic escalation flagging, still with no orchestration framework until the if/elif routing genuinely stops being legible. Recent improvement: the escalation trigger (severity_flag) went from 25% specificity to a measured 100%/100% recall-and-specificity, via a properly balanced ground-truth set instead of a token 1-positive-case sample.

Bachelor in Data Science, University of Valencia. 3+ years of professional experience:

• IVIRMA Global — ML demand forecasting in Python, NLP/sentiment analysis on unstructured text, end-to-end ETL and ML pipelines with SQL, PySpark, and Azure Databricks; model evaluation, A/B testing, and hyperparameter tuning with Azure ML; Power BI/Tableau dashboards for stakeholder reporting.
• SDG Group — ETL workflow development and datamart design/validation for a banking client, using SQL and PowerCenter.
• University of Valencia (NLP project) — NLP-based data anonymization and text-preprocessing pipelines, with automated ETL and data-governance workflows for privacy-sensitive data.

Comfortable across the full loop of a data problem — from analysis and modeling to shipping a result as a service — with the AI-agent work above as the current, self-directed focus.

• GitHub: you're already here — @serpeigd
• LinkedIn: sergio-peigneux