A curated set of AI/ML builds presented like product artifacts: problem statement, architecture, constraints, and tradeoffs.
Each entry is designed to be understandable in ~10 seconds—click, run, and see what it does.
ProtagorasMulti‑LLMOrchestration
Protagoras (v1.5.0)
A multi‑participant debate forum that can mix providers (OpenAI / Gemini / Anthropic / xAI). It keeps a persistent transcript,
runs multiple rounds, and ends with a moderator synthesis. It also renders an optics‑first similarity matrix computed from the
last round of responses.
How to use (10 seconds)
- Type a decision question.
- Click Start New Debate (or Send Follow‑up).
- Skim the synthesis and similarity matrix.
PromptingWorkflowCoaching
Prompt Builder v2.17
A prompt engineering workbench: build prompts from explicit components (role, task, context, constraints, output format),
run them live, and iterate using “Prompt Coach” critique. The goal is to make prompts more like reusable software than ad‑hoc chat.
How to use (10 seconds)
- Select modules and click Build Prompt.
- Click Run Prompt to test.
- Use Coach feedback to iterate.
Computer VisionMNISTIn‑browser
Using Machine Learning for Recognition of Handwritten Digits
A MNIST digit classifier you can run in the browser: draw a digit (0–9) on a canvas and the model predicts.
MNIST is a classic dataset of handwritten digits often used to teach and benchmark image classification.
FreeOrgDecision OS
FreeOrg Web Tool (concept)
Decision‑making workflows with structured debates, reusable meeting formats, and outputs tuned for action.
This is the conceptual layer behind the Protagoras‑style interface: turn fuzzy questions into structured dialogue,
capture assumptions, and converge on a concrete decision.
Demo intentionally omitted (concept‑only)
Coming soonAgentsQuant
Agentic AI Trading Labs
Build agent‑driven trading desks and laboratories working 24/7 to seek out investment opportunities and trade strategies.
Agents will whiteboard ideas, backtest, and iterate on strategies across assets and derivatives with conservative risk
management and compliance baked into the workflow.
The emphasis is a falsifiable research loop: pre‑register hypotheses, run kill‑tests, log assumptions, and promote only
strategies that survive robustness checks. Outputs are “decision artifacts” (what, why, risks, next steps), not vibes.
Write-up in progress
Coming soonNLPLLMs
Natural Language Processing workshop
A hands‑on collection of demos spanning transformers, embeddings, RAG, chunking, and practical fine‑tuning.
Covers fundamentals (NER, sentiment, POS, coreference) and modern production topics like evaluation, long‑context strategies,
GPU efficiency, mixture‑of‑experts, and reasoning‑oriented prompting.
Each lesson is paired with a runnable artifact and a “how to evaluate” checklist—so the reader can verify behavior quickly.
Write-up in progress