"If it isn’t reproducible, it isn’t engineering."
Local AI is back—quietly, and for practical reasons
Local-first AI isn’t a trend. It’s a response to real constraints: cost, latency, privacy, reliability, and the need to understand what your system is doing—step by step—when the workflow gets complex.
The problem is that most “local AI” guides stop at the fun part: running a model once. The hard part starts right after: building workflows that you can run again tomorrow, debug next week, and extend next month without the whole stack turning fragile.
A lab, not a dump folder
This repository is organized as a portfolio-grade lab: a clear roadmap plus a sequence of sessions. Each session adds a real capability—an API contract, a UI, a multi-step flow, structured outputs, observability, or an external bridge— while keeping the experience grounded in engineering discipline.
The goal is simple: make progress visible. Make experiments repeatable. Make complexity earn its place.
What you’ll find inside (Sessions 1 → 6)
Session 1 — A stable local baseline
A clean starting point: a simple API surface that proves local inference is running reliably before anything “agentic” is added. Think: predictable inputs, predictable outputs, and a foundation you can build on without guesswork.
Session 2 — From chat to an actual agent runtime
Tools, memory, and a lightweight UI turn a plain endpoint into something that behaves like a system. This is where “agentic” becomes concrete: routing, state, and controlled execution.
Session 3 — Multi-agent orchestration with roles
A crew-style pipeline with distinct roles (research, analysis, writing, review) designed to make reasoning more inspectable and outputs more deliberate. The emphasis is orchestration and evaluation patterns—how multi-step systems behave when roles are separated.
Session 3.4 — Persistent context, quant signals, charts, and PDF output
This is the “systems upgrade”: persistent retrieval for reusable context, quantitative signals and visualizations, and a workflow that produces a concrete artifact (a PDF report) instead of only a chat reply.
Session 4 — A web Research Operator (no private corpus)
A research-focused operator designed for the real world: it can run multi-step web research, present results in a UI, and support observability so you can see what happened—not just what it produced.
Session 5 — Research Operator with dynamic agent teams
Same research objective, different orchestration style. This session explores a more dynamic “team creation” approach: spinning up roles on demand, managing turn-taking, and returning structured summaries about what the system did.
Session 5.1 — A crew-based variant with strict contracts (and what it teaches)
A focused comparison that stresses structured communication across steps. It highlights a practical lesson: strict contracts can improve clarity, but they can also introduce fragility and overhead if the pipeline isn’t designed carefully.
Session 6 — An MCP Trading Desk bridge (MT5 + agentic pipeline)
A systems-level capstone: a workflow that accepts market snapshots from MetaTrader 5, runs staged evaluation, uses configurable guardrails, and exposes an MCP interface alongside web/API routes. It’s a practical example of “agentic workflows meeting external, real-time systems.”
Why this is worth opening
Because it doesn’t ask you to believe. It asks you to run.
You can observe how endpoints are shaped, how memory and tools change behavior, how orchestration strategies differ, where structured output becomes brittle, and why observability matters the moment your workflow has more than one step.
Explore it
If you’d like to browse the full lab and follow the track end-to-end, start here: