Product Case Study

AI Investment Manager

A real, actively-developed research platform that earns the right to use AI one gate at a time — evidence first, human approval always, automation last.

Most AI investing tools lead with a claim about smarter picks. This one leads with a capability-gate model: research and evidence today, conservative human-approved recommendations and paper trading in progress, real money and automation locked until they're earned.

Product

Self-hosted, local-first research platform

Scale

468+ commits, actively developed

Stack

FastAPI + Postgres + Streamlit

AI Status

Zero LLM calls in product to date

Real, In-Progress Product — This page recreates the real Command Center interface from an actively-developed personal product. The gate statuses shown (delivered / started / locked) reflect the real product's actual current state. No AI/LLM-generated content exists in the product yet at any gate; all recommendation and workflow logic shown is deterministic and human-reviewed.

Try It

The real Command Center

This is the actual start-here console from the real product — a next-best-action engine and a workflow-status grid, not a stock picker. Click any section to see what it actually covers.

Next Workflow Action

medium

Review signals awaiting evidence

AAPL and NVDA each have a grounded signal ready for human review before it can move toward a recommendation packet.

Opens: Review Pipeline

This is workflow routing only. It does not recommend investments, rank tickers, score, allocate, or trade.

Capability Gates — What's Real Today

Gate 1

Evidence & Research Foundation

Delivered

Provenance, signals, and the human review workflow.

Gate 2

Investment Cockpit & Manual Portfolio Context

Delivered

Watched tickers and manual portfolio context — organization, not advice.

Gate 3

Explainable Human-Approved Recommendations

Started

Conservative recommendation packets — watch / research more / avoid — human approval required.

Gate 4

Paper-Money Simulation

Started

Fake-money simulated orders and positions, seeded only from approved packets.

Gate 5

Real-Money Broker Integration With Approval

Locked

Broker read-only sync first, then user-approved trade tickets with risk controls.

Gate 6

Guardrailed Automation

Locked

Limited automated execution within strict guardrails, disabled by default.

Case Study

The story behind the system

The Insight

AI investing tools compete on hype, not trust

Most AI investing products lead with a claim — smarter picks, better signals — that's easy to copy and hard to prove. The differentiated position isn't a smarter model; it's a system that can show exactly what it knew, where it learned it, and why it reached a conclusion.

The Product Thesis

Evidence before recommendations, paper trades before real money

Every capability unlocks behind an explicit, named gate: research and evidence first, then conservative human-approved recommendations, then fake-money simulation, then — only later, and only with approval — real trading. AI is introduced deliberately, once earlier gates have proven trustworthy, not on day one.

The Outcome

A working evidence-to-decision pipeline, built without AI, by design

Source tracking, signal review, human-approved recommendation packets, and fake-money paper simulation are built and functioning end to end — entirely deterministic, zero LLM calls, with a human required to approve every consequential step.

My Role

Sole builder — product, architecture, and the gating model

Product Leadership

  • Designed the capability-gate model itself — six named gates, each unlocking only when explicitly started, so AI and automation are earned rather than assumed
  • Wrote the public positioning strategy: lead with transparency and evidence, not with claims of beating the market
  • Defined the conservative recommendation action set (watch / research more / avoid for now / paper-track candidate) — deliberately excluding buy/sell, price targets, or allocation until much later gates
  • Set the human-approval requirement as a non-negotiable boundary at every gate, including the ones already built

Technical Direction

  • Backend: FastAPI + Postgres, with an append-only audit trail across sources, signals, packets, and paper orders
  • Dashboard: a modularized Streamlit app (shared API/state/UI helpers, extracted page renderers) covering 30+ workflow pages
  • Built the Command Center console: a deterministic next-best-action engine aggregating status across every workflow area
  • Built the recommendation-packet and paper-simulation-handoff state machines, including a 9-section approval checklist and an 11-state paper-handoff eligibility engine

Outcome

What's real today

  • Gate 1 (Evidence & Research Foundation) and Gate 2 (Investment Cockpit & Manual Portfolio Context) are fully delivered.
  • Gate 3 (Explainable Human-Approved Recommendations) and Gate 4 (Paper-Money Simulation) have started, in conservative, deterministic form.
  • Zero AI or LLM calls exist in the product to date, at any gate — every signal, score, and recommendation is deterministic and evidence-linked.
  • Every consequential action, from a recommendation packet to a paper trade, requires explicit human approval.

What's Next

  • Gate 3's AI-assisted explanation layer — synthesizing evidence into plain-English rationale, still human-approved, never auto-acting
  • Gate 5: real-money broker integration, starting with read-only sync before any user-approved trade ticket
  • Gate 6: guardrailed automation, disabled by default, only after real-money gates prove trustworthy
  • Expanding the paper-trading trust ladder with performance tracking and prediction-accuracy scoring

Why This Matters

The hardest part of building an AI investment manager isn't the AI. It's proving, one gate at a time, that it deserves to be trusted with your money.