I'm not the genius.
I'm the convergence point.
Since late 2024, Claude and Gemini have been fully integrated into every stage of my production work — not as occasional accelerators, but as the default first step on every problem. AI agents surface patterns across datasets too large to audit manually. Open source repos distill thousands of engineering-hours into proven, battle-tested architecture. Developers translate requirements into feasibility signals. Domain knowledge filters what matters from noise: which regulatory requirement is genuinely binding, which "best practice" was written for a different context, which user behaviour is signal versus edge case.
My job is to sit at the convergence of all four — and produce the ranked solution set. Not the most impressive one. Not the most technically elegant one. The one that ships, withstands regulatory scrutiny, and doesn't require rebuilding when the next update lands. This is not a methodology I'm pitching. It's a production protocol I run every day. Five demos follow — each one documents a live decision from real work. Every decision traceable back to a signal, a source, and a reason.
for structured analysis
for long-document research
every project, every day
The Signal Stack
Every problem I work on gets run through the same convergence model — and has done since late 2024. Four signal sources feed into one terminal. The terminal produces ranked solutions. The highest-ROI solution ships. What follows is not a portfolio of impressive-looking interfaces — it's a portfolio of real decisions made on real production work, each one traceable back to a signal, a source, and a reason.
ROI Solution Explorer
Every problem has more than one solution. Most teams pick the first one that sounds reasonable, or the one the most senior person in the room proposed. The Terminal Model doesn't do that. The scoring framework here was directly informed by studying ai-hedge-fund's multi-agent analysis structure — 13 agents scoring the same signal across independent dimensions, then surfacing disagreement rather than averaging it away. The same principle applied to design decisions: generate multiple solution pathways, score each across four ROI dimensions independently, and document why the rejected paths were rejected — not just what was chosen.
Three real problems below — all drawn from actual ACY Securities work. The solutions are ranked, not just listed. The reasoning is shown, not summarized. This is the decision protocol in its raw form, before it becomes a polished deliverable.
OpenStock: Consumer Grade → Institutional Grade
OpenStock is a well-engineered open source platform. The engineering is solid. The UX serves retail investors well. It was the first candidate I evaluated when ACY needed institutional stock data visualization — and it was rejected. Not because the code was bad, but because its AGPL licence requires open-sourcing any institutional customisations, which would expose competitive IP.
That's an open source distillation decision: study the architecture, apply the UX learnings, integrate TradingView instead. Toggle below to see the exact institutional layer that OpenStock's codebase informed — even though OpenStock itself wasn't shipped.
- Clean price display — reduces cognitive load for self-directed retail investors
- Full-width chart — maximum data visibility for casual analysis
- No compliance layer — retail platforms aren't legally required to show it
- ① Compliance bar persists — regulatory status cannot be buried in a menu at the point of trade execution
- ② Position calculator at decision point — risk data belongs where the trader is looking, not in settings
- ③ VaR Floor price line — maps risk model output directly onto price chart, connecting analysis to action
- ④ Audit trail auto-generates — MiFID II requires this. Designing it in is cheaper than building it later.
AI Signal Trust Layer
"87% confident." That's a lie. Not intentionally — but structurally. Kronos (AAAI 2026) doesn't output a single confidence percentage. It outputs a probability distribution over thousands of future candlestick paths. Reducing that to one number misrepresents the model — and in institutional finance, that misrepresentation is a compliance risk, not just a UX flaw.
ai-hedge-fund's multi-agent architecture adds a second layer: 13 investor personas often reach opposite conclusions. That disagreement is the signal — not a problem to hide with an averaged consensus score. Three approaches below. One is structurally wrong, one is technically accurate but practically inaccessible, one ships with MiFID II built in.
KYC Drop-off Optimization
73% of users abandoned ACY Securities' KYC onboarding. The instinct is to rebuild the whole flow. That would have taken six weeks and missed the ASIC compliance deadline. The Terminal Model applied a different question first: where exactly are they leaving?
I fed Mixpanel event data into Claude with a structured prompt: identify the two highest-drop steps, classify abandonment by device type and user segment, surface the pattern. Two days, not two weeks. AI identified two specific screens causing 60% of abandonment. I redesigned those two screens — not the whole flow. Shipped in three days. 73% → 45% drop-off, measured over 90 days post-launch.
Identity Verification
Complete all fields below to verify your identity
By submitting this information, I confirm that the details provided are true and accurate to the best of my knowledge. I understand that providing false information may result in the termination of my account and potential legal consequences under applicable laws and regulations including but not limited to the Anti-Money Laundering and Counter-Terrorism Financing Act 2006 (Cth), the Corporations Act 2001 (Cth), and relevant ASIC regulations.
Verify your identity
Step 2 of 4 · Takes about 3 minutes
What it actually looks like
Demo 04 states "I fed Mixpanel event data into Claude with a structured prompt." Below is a reconstructed excerpt from that session — the actual input structure and the output that scoped a 6-week rebuild down to a 3-day sprint. The numbers on the demo page came from this exchange.
Context: ACY Securities KYC onboarding funnel. ASIC-regulated broker. Primary market Asia-Pacific.
Dataset: Mixpanel event export · 12,847 sessions · Feb–Mar 2024 Task: 1. Identify the two steps with highest abandonment — session counts and device breakdown (mobile vs desktop) 2. Classify by user segment (new vs returning) 3. Surface the pattern — form fatigue, upload friction, or cognitive load? 4. Tell me what NOT to redesign. Output: ranked findings, step names, percentages, recommended scope. Constraint: 3 days before ASIC compliance window closes.
Analysis of 12,847 KYC sessions — drop-off by step: ───────────────────────────────────────────────────── Step 3 · Identity Verification 41% abandon ← highest Step 5 · Risk Assessment 19% abandon ← second Step 2 · Personal Info 8% abandon (baseline — acceptable) Step 4 · Agreement 5% abandon (performing well) Top 2 steps account for 60% of total abandonment. STEP 3 PATTERN — Identity Verification: Mobile: 58% abandon at document upload prompt → upload friction Desktop: 23% abandon → form field count (6 fields, one screen) New vs returning: no significant difference → structural issue, not familiarity STEP 5 PATTERN — Risk Assessment: Uniform across device types (±2%) → question complexity, not UI New users: 31% abandon · Returning: 8% → first-time cognitive overhead RECOMMENDED SCOPE: Redesign Steps 3 and 5 only. DO NOT REDESIGN: Steps 1, 2, 4. Step 2 at 8% is within acceptable range for ASIC-compliant identity capture. Step 4 at 5% is performing well — changes risk introducing new drop-off. Full flow rebuild is unnecessary and exceeds the 3-day constraint.
Agent-Native vs Human-Readable UI
AI-Trader makes something explicit that most platforms pretend doesn't exist: "Just like humans have their trading platforms, AI agents need their own." Its platform serves both human traders and AI agents simultaneously — both consuming the same signal feed, both executing via the same infrastructure.
The design problem this creates is real and under-discussed: designing one interface to serve both cognitive architectures forces compromises that degrade both. The human needs visual hierarchy, narrative framing, and social proof to reduce uncertainty. The agent needs none of that — it needs numeric precision, machine-parseable structure, and a direct execution path. Toggle below to see exactly what changes, and why.
{ "signal_id": "sig_kaa_20260416_001", "provider_id": "agent_kronos_alpha_7f3a", "asset": "AAPL", "direction": "BUY", "entry_price": 182.50, "confidence": ① 0.87, "position_size_pct": ④ 0.04, "stop_loss": 174.80, "take_profit": 198.20, "risk_reward": 2.18, "track_record_30d": ② { "win_rate": 0.71, "sharpe": 1.43 }, "expires_at": "2026-04-16T20:00:00Z" } // ③ No social fields. No community signals. Agents optimize for Sharpe, not consensus. // Execution call — agent recalculates allocation against its own portfolio: POST /api/claw/copytrade { "signal_id": "sig_kaa_20260416_001", "max_risk_pct": 0.02, "broker": "interactive_brokers" }
The GitHub Network I Distilled From
These repos aren't projects I built. They're signal sources I studied and distilled — each one feeding a specific insight into a specific design decision. The connection between research and shipped work is documented below.
All Five Patterns. One Terminal.
Five demos, each isolating one aspect of the Terminal Model. The TradeX Institutional Terminal is what happens when all five converge on a single product — a fully-functional institutional trading simulation with live charts, FIX 4.4 protocol engine, Design Mode annotations tracing 19 decisions back to their signal source, and a live Agent API toggle on the Kronos panel showing exactly how the same data transforms when the consumer switches from human to machine.
I'm not a genius. That's the point.
Since late 2024, I haven't worked on a single significant problem without running it through Claude first. Not as a shortcut. As a protocol: define the problem precisely, surface the signal sources, generate multiple solution pathways, score by ROI across four dimensions. By 2025 Gemini was part of the chain. By 2026 this is just how I work — the same way a senior engineer reaches for the debugger before guessing, I reach for the model before opinionating.
The next generation of institutional products won't be built by the smartest person in the room. They'll be built by the designer who can precisely define the problem, orchestrate the right signal sources, synthesise the inputs, and output the highest-ROI solution — repeatably, under regulatory constraint, with a team that doesn't all have to be geniuses either.
Five demos on this page. Five different problem classes. One protocol applied consistently: generate multiple solution pathways, score by ROI across four dimensions, document the rejected paths, ship the chosen one with the reasoning intact. Demo 01 shows the decision framework directly. Demo 02 shows open source research turned into institutional product. Demo 03 shows where AI output requires human judgment by design — and how to make that handoff compliant. Demo 04 shows how to use AI to find the right scope before touching the design. Demo 05 shows a problem most designers haven't encountered yet: what happens when the user is an AI agent, not a human.
That last one matters more than it currently seems. AI-Trader, MiroFish, ai-hedge-fund — the signal across all three is consistent: agent-to-agent systems are already in production, and the UX decisions being made now about how those systems present and consume data will shape institutional finance interfaces for the next decade. The designer who understands both cognitive architectures — human and machine — and has been shipping work with both since 2024, will be indispensable in that transition.
The failure mode to watch: AI confident on things it shouldn't be. Post-training-cutoff regulatory changes, specific financial figures, client relationship context, institutional political dynamics — none of these should be delegated. The AGPL rejection in Demo 02 wasn't a Claude call. The MiFID II compliance architecture in Demo 03 was validated against primary ESMA text, not taken from AI output. The Terminal Model doesn't replace domain knowledge — it processes faster so domain knowledge has more precise inputs to work with.
If you're building products where the margin for error is regulatory
Five years in institutional finance. 100K+ traders on live systems. 40+ jurisdictions. Eight regulatory updates absorbed without structural rework. And since late 2024 — every one of those decisions made with AI fully integrated into the production chain. I'd like to show you what the Terminal Model looks like applied to your product.