Skip to content
@genial-labs-ai

Genial Labs

AI

Genial Labs

Machine learning workshops taught from working code. You clone the repository and run the labs; the slides are there to explain what you just saw.

The bias throughout is towards evaluation. A model is the easy half of a system. The other half — the harness that calls it, the tests that hold it to account, the gate that stops a regression reaching production — is where the failures live, so that is where the labs spend their time.

Workshops

Four days. Workshop site · agent-eval-ci

One running example: Stockroom, an inventory and order-support agent with five tools, a 50-case golden set, and four deliberately seeded weaknesses — an ambiguous tool description, an oversized tool payload, an unbounded retry, an unguarded prompt injection. Each is a feature flag, so you switch it on, measure what it costs, then fix it in the harness. The ambiguous description alone drops tool-selection accuracy from 1.00 to 0.64 in the deterministic mock run while answers can still look right, which is the point of the workshop.

Along the way: deterministic trajectory metrics, LLM judges calibrated against 32 human-labelled cases and probed for position, verbosity and self-preference bias, OpenTelemetry traces, tools served over MCP, a red-team suite, and a GitHub Actions gate that fails the pull request when a metric slips against the committed baseline. All of it runs offline against a deterministic fake model and fake judge, so no cloud account is needed. Amazon Bedrock is opt-in; live CI uses GitHub OIDC with no long-lived keys. A live run caps its own tokens, steps and wall-clock time, and reports what it used.

Five days. Workshop site · Schedule

Fifteen modules from n-grams to agents: text as data, word vectors, sequence models, attention, a transformer written from scratch, pretraining on the Hugging Face stack, LoRA fine-tuning, preference learning and RLHF, calibration, retrieval-augmented generation, and a capstone that fills most of the last day in pairs. Modules 1–14 pair a short lecture with a Colab notebook; Module 15 is the hands-on capstone. API keys are optional: labs provide free open-model paths or, for calibrated decisions, a small decision model trained in the lab.

Still being hardened for the Colab runtime — the readiness page states exactly which labs have run end to end and where. Development happens at project-delphi/nlp-llms; the copy here is a mirror.

One day. Workshop site · Slides EN · Diapositivas ES

From "I know matrices" to manipulating, solving and factorizing matrices and tensors, in Colab notebooks with knowledge checks and live deep dives between them. It assumes one chapter of linear algebra and no tensor theory whatsoever; every term is defined where it first appears. Taught in English, with the slides, handbook and references also in Spanish. Development and the live site are at project-delphi/tensors-workshop; the copy here is a read-only mirror.

How we work

  • No required paid model APIs. Workshop labs provide paths without paid model API keys. Stockroom runs offline; the NLP labs use open-model downloads or locally trained models. Commercial model calls are opt-in, with their usage and cost limits stated.
  • Reproducible numbers. Figures quoted in a lecture come out of the repository that ships with it, in a mode anyone can rerun. Post-mortems are labelled composite or linked to a source. We do not invent incidents, companies or statistics to make a point land.
  • Agent = Model + Harness. Every claim about a failure mode arrives with the test or notebook cell that demonstrates it.
  • Yours to teach from. Fork it, teach it, send a pull request. Each repository carries its own contribution guide and licence.

genial-labs.com · ravi@genial-labs.com

Pinned Loading

  1. system_agent_harness_aws system_agent_harness_aws Public

    Evaluating Autonomous Agents: Systems, Harnesses & AWS Production CI/CD — a 4-day workshop around the Stockroom agent

    Python

  2. nlp-llms nlp-llms Public

    From Traditional NLP to Modern LLMs: a five-day workshop by Genial Labs (mirror of project-delphi/nlp-llms)

    Jupyter Notebook

  3. tensors-workshop tensors-workshop Public

    Mirror of project-delphi/tensors-workshop — a bilingual (EN/ES) Quarto site for a one-day tensors workshop. Read-only; development and the live site are at the canonical repo.

    Jupyter Notebook

Repositories

Showing 4 of 4 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…