AI agents that work where fintech teams work.
Train it on company knowledge. Connect it to enterprise tools. Deploy it inside the products teams already use. Not a chatbot. An intelligent layer that actually does things.
I designed the full system — research, architecture, interaction model, builder experience, and end-user interface across four fintech roles.
How do you design an AI system that financial professionals will trust with real money decisions?
Every agent action must be auditable. Compliance is non-negotiable.
Traders need sub-second responses. 3 seconds = missed opportunity.
One wrong margin number could cost millions. Accuracy beats speed.
Financial teams aren't slow because they're bad at their jobs. They're slow because their tools don't talk to each other and their knowledge is buried in documents nobody reads.
Finding a margin policy takes 20 minutes of searching through a 240-page manual. Multiply by 15 queries per day.
CRM → OMS → compliance rules → Jira. Four tools, four logins, zero integration between any of them.
Reconciliation, compliance checks, and reports are rule-based — but they consume hours that should go toward judgment work.
Conceptual research based on 12 published fintech workflow studies, SEC enforcement data, and enterprise AI deployment post-mortems. Personas are synthesized archetypes — not direct interviews.
Manages 40+ positions daily. Needs market data, risk metrics, and execution speed in one view.
Reviews flagged transactions, prepares audit reports. Citation accuracy is everything.
Monitors portfolio risk across funds. Needs to spot problems before they escalate.
Handles settlement, reconciliation, and exception management.
Trust is the product, not intelligence. An agent that's right 95% but can't show its work is worse than one at 90% that cites everything.
Users don't distrust AI — they distrust outputs they can't verify quickly. If citation + verification takes 30 seconds, adoption follows. If 5 minutes, they'll search the PDF themselves.
The high-value work is connecting systems, not answering questions. An agent that orchestrates across CRM + risk + Jira is 10x more valuable than one that chats.
Different roles need fundamentally different agent personalities. Same brain. Different error tolerances.
The builder (admin) is the forgotten user. Their frustration directly determines agent quality.
"I don't need AI to be smart. I need it to be right, and I need to prove it was right when auditors come knocking."
— Synthesized from compliance analyst patternsThis insight became the foundation for every design decision.
FinPilot has two sides. The builder side — upload knowledge, connect tools, set permissions, test. The user side — the agent deployed inside the products teams already use.
Same trained agent, different face depending on who's using it. Traders get speed. Compliance gets citations. Ops gets checklists.
Research showed teams won't adopt another tool. The agent must live inside existing workflows.
A trader needs action buttons. Compliance needs audit trails. One interface can't serve both.
Train on company documents using RAG. Upload → chunk → embed → retrieve.
Connect OMS, risk engine, CRM, compliance DB. Agent reasons, calls tools, synthesizes.
Embedded copilot, proactive alerts, standalone workspace, backend automation.
Step by step. Where the agent helps. Where the human decides. Where it breaks.
FinPilot detects a trade pattern matching wash sale criteria. Marcus sees a notification — he didn't search for it.
One click opens a side panel with both trades, time gap, client details, and the relevant regulation — all pulled automatically.
OMS.get_trade_history(client: 4521, ticker: MSFT, 30d)ComplianceDB.get_rule(wash_sale_irs_1091)CRM.get_client_profile(4521)Agent summarizes the case. Every claim links to its source. Confidence: 87% — the 13% gap is flagged explicitly.
Sell was in taxable account, repurchase in IRA. The agent doesn't guess — it flags: "Cross-account rules apply but enforcement varies. Recommend manual review."
The agent states what it doesn't know. Provides the regulation text. Flags for human judgment. This is where Marcus's expertise matters.
After reviewing cross-account guidance, he confirms: wash sale violation. Agent executes the mechanical follow-up — not the compliance decision.
Audit record, Jira ticket, compliance database entry — all generated automatically with full citation trail.
ComplianceDB.log_investigation(WS-2025-0342)Jira.create_ticket(client_notification, medium)OMS.flag_trade(7842, wash_sale_confirmed)Same investigation. Fraction of the time. Because the agent did the research, and the human did the judgment.
The agent never made the compliance decision. It gathered evidence, flagged ambiguity, and documented the outcome. The human made the call.
Compliance investigation — before and after. Estimated from workflow simulation.
Scan monitoring report for flags
Log into OMS, search trade history
Open compliance PDF, search wash sale rules
Cross-reference trade dates manually
Log into CRM for client details
Write summary in Word
Create Jira ticket manually
Log findings in compliance DB
Proactive alert with pre-compiled context
Agent pulls trade history, regulation, client data
Analysis with source citations + confidence
Ambiguity flagged — human makes the call
Audit record, Jira ticket, compliance log auto-generated
One wrong margin number = millions in liability. The trust layer isn't a feature. It's the product.
Every claim links to document, section, and version. Click to see original text. Not optional — hard requirement for compliance.
90%+ = green badge. Below 80% = explicit uncertainty flag. Below 60% = refuses to answer definitively.
Expand any response to see: which tools called, which docs retrieved, how the answer was synthesized. Makes the agent's thinking auditable.
Doesn't know? Says so. Tool fails? Clear error + retry. Outdated source? Flags the last-updated date. No silent failures.
Auto-execute: Read-only queries, data lookups, reports.
Confirm to proceed: Trade modifications, escalations. Agent suggests, human approves.
Human only: Policy changes, regulatory submissions, compliance decisions.
Every interaction logged: query, tools called, sources retrieved, response, user action. Immutable and timestamped.
Audit logs retained 7 years per SEC Rule 17a-4. Written to immutable store, not app cache.
Company knowledge → searchable AI memory.
Documents get broken into chunks, converted to vectors, stored in a specialized database. When someone asks a question, the system finds the most relevant chunks and gives them to the AI as context. No retraining needed.
Fine-tuning = retrain the whole model. Expensive, slow, hard to update. RAG = upload a new doc, it's searchable in minutes.
RAG for factual, document-grounded answers. Fine-tuning for domain-specific reasoning. FinPilot uses RAG primarily.
Compliance manuals, trading policies, SOPs, filings.
Extract clean text. Handle PDF layouts, tables, images.
Semantic splits at topic boundaries. Size matters.
Convert chunks to vectors. Similar meaning = similar numbers.
Vector DB + metadata: source, page, section, date.
Query → embed → find nearest chunks → feed to LLM.
Tools that let the agent act, not just talk.
Trade status, orders, positions
Stress tests, exposure, VaR
Account info, history, records
Rules, violations, audit logs
Prices, VIX, historical
Slack, email, Jira
Think → Plan → Execute → Observe → Respond. If insufficient, loop again.
Four ways the agent reaches users.
Side panel. Context-aware. Knows what screen you're on. Ask a question, get an answer without leaving your workflow.
Full-screen for complex multi-step conversations. Scenarios, comparisons, report generation.
Doesn't wait. Watches metrics, pushes alerts with context and recommended actions.
No UI. Nightly reconciliation, daily compliance checks. Only surfaces exceptions.
Most AI case studies skip this. But the person uploading documents and debugging wrong outputs is the most underserved user.
"I uploaded our entire compliance library and the agent still gives wrong answers about margin calls. No idea which document is causing it."
— Synthesized from enterprise AI admin frustrations340+ files. Inconsistent formatting, scanned pages, watermarks, broken tables.
Admin doesn't need vector math. They need to verify the system understood their documents.
Agent gives wrong margin answer. Most tools show the wrong answer but no way to trace why.
Debugging panel shows retrieval chain. Wrong chunks retrieved = document pollution. Right chunks but wrong answer = model reasoning issue. UI flags the distinction.
"Agent says it doesn't know our options SOP, but I uploaded it last week."
Saved test suite. Automatic accuracy scoring by topic. Hit 90%+ → promote to production.
Target admin is PM/ops, not ML engineer. Replaced with 3 presets: Detailed / Balanced / Broad — with visual preview of how each splits a sample document.
Styled low-fidelity. Layout and interaction patterns — not visual polish.
The home base for agent administrators. Shows agent health, usage metrics, knowledge base status, and quick actions. Designed for a PM or ops lead who manages the agent.
Tool health cards show latency numbers, not just green/red status. In financial systems, a "connected" tool with 320ms latency is functionally degraded — traders notice. Showing the number lets admins act before users complain.
Where admins upload, manage, and monitor the company documents that power the agent's knowledge. Shows upload interface, document list, chunk preview, and embedding status.
Admins need to verify the agent's knowledge is correct before deployment. The chunk preview lets them see exactly what text the agent will retrieve for a given topic — catching bad parsing, broken tables, or garbled PDFs before they cause wrong answers in production.
Where admins test agent responses before deployment. Shows the conversation, source citations, tool calls made, and confidence scoring — so teams can verify accuracy before going live.
Early concepts hid the reasoning trace behind a "Show details" toggle. We surfaced it permanently because admins testing the agent need to see not just what it said, but why — which sources it used, which tools it called, and how long each took. Hidden details don't get checked. Visible details build trust before deployment.
The agent appears as a collapsible side panel within the trader's existing terminal. It has full context of the current view and can execute actions directly. Speed-optimized with one-click confirmations.
Traders operate in milliseconds. Typing "execute the collar strategy" and waiting for a response costs time. The one-click "Execute collar →" button appears directly in the agent's recommendation — reducing the action to a single tap. The agent suggests, the trader confirms with muscle memory, not conversation.
The agent monitors risk metrics continuously and pushes alerts when thresholds are breached. Shows the proactive notification banner, the risk dashboard context, and suggested actions — the agent reaches out before the user asks.
Early designs had the agent recommend a single "best" action. Rejected because risk management is judgment-heavy — the "best" option depends on market outlook, client relationships, and risk appetite that the agent can't fully assess. Three ranked options let the risk manager apply their expertise to the agent's analysis. The agent does the math; the human chooses the strategy.
A dedicated, full-screen environment for deep multi-step analysis. Used when embedded panels aren't enough — complex report generation, multi-system queries, and iterative research with the agent.
Lives inside existing workflows. Never pulls users away.
Show sources, reasoning, confidence. If you can't cite it, don't say it.
Suggestions first. Earn trust. Then automate.
Same brain, different behavior. Different error tolerances.
Doesn't know? Says so. Tool fails? Shows why. No silent failures.
Do users trust confidence scores? Show same response with/without confidence to compliance analysts. Measure manual verification rate.
Will compliance teams accept AI-generated logs? Present auto-generated investigation records. Ask: "Would you submit this to auditors?"
Is the 80% confidence threshold right? Run 200 queries at varying confidence. Calibrate against analyst error tolerance.
Prototype compliance flow. Test with 5 analysts. Measure: completion time, verification steps, trust (1-5).
Builder prototype. Can 3 ops admins identify and fix a wrong answer within 10 minutes?
A/B test: citations-only vs citations + confidence vs citations + reasoning trace. Lowest manual verification wins.
Before/after estimates (~2h → ~12m) are from workflow simulation. Live numbers would differ — likely better for routine, worse for edge cases.
Research, architecture, interaction design, trust engineering — and the product decisions that make AI systems safe enough for financial professionals to rely on.
I'm looking for roles where I can design AI products at this depth.
UX / Product Design · 6+ years enterprise platforms