Evidence layer
Source hierarchy
SEC filings, issuer announcements and official sources establish company facts. Reputable current market data provides prices and instrument context.
Independent project / technical case study
Agentic
Investment Lab
A governed AI research system built to discover, challenge, and document investment ideas, with explicit boundaries between autonomous analysis and human decisions.
Explore the sanitized architecture, agent responsibilities, decision logic, operating cadence, and risk controls behind this private experiment.
Explore the architecture ↓Try the decision treeResearch and execution are separateDecision-ready does not mean trade-ready
Human approval requiredNo unattended brokerage actions
Evidence and audit firstFresh sources and durable reasoning
01 / System architecture
Specialized workflows, connected by durable state
Instead of a single agent making every decision, independent workflow stages collect evidence, challenge conclusions and enforce separate gates.
DiscoverBroad-market candidates
ValidateSources, catalyst, liquidity
ChallengePrimary / red-team cases
AdjudicateInvestment Committee
EscalatePrivate verification + human
Reasoning layer
Independent review
Each candidate is evaluated for catalysts, downside, financing, liquidity and alternatives, with explicit counterarguments.
Persistence layer
Decision records
Notion-style records retain research, lifecycle status, next-review dates and audit history. Stored state never substitutes for live quotes or brokerage facts.
02 / Interactive workflow explorer
Select a workflow to inspect its boundaries
The workflows have different permissions, research depth and escalation behavior. Explore the sanitized responsibilities below.
Independent opportunity discovery
Search across eligible markets without anchoring on existing holdings. Apply quality and liquidity filters, verify a credible catalyst, and compare opportunities with the option of holding cash.
- Input: public sources and current market context
- Output: ranked research candidates, rejected ideas and evidence gaps
- Stop condition: insufficient evidence, questionable liquidity or promotional risk
Research underwriting
Build thesis, invalidation, catalyst timeline, upside/downside cases and instrument suitability. Update observations only when sources support a change.
- Input: filings, issuer guidance, verified market and option details
- Output: documented thesis, exit plan, dated evidence
- Stop condition: unresolved material contradictions or stale information
Investment Committee adjudication
Evaluate a primary case, independent red-team critique and alternatives before determining whether research is actionable, watchlisted, rejected or requires re-underwriting.
- Input: complete current research and counter-thesis
- Output: documented conclusion and next review
- Boundary: research approval is not execution permission
Exception-oriented market watch
Lightweight scheduled checks flag new filings, catalysts, material events and thesis changes. Routine no-change runs produce status without creating unnecessary research records.
- Input: current public facts and tracked research state
- Output: heartbeat, material exceptions or review request
- Boundary: no unattended private brokerage reconciliation
Interactive execution-readiness gate
Before a proposed trade can advance, current private account state, concentration, drawdown, available funds, orders, instrument access and the final broker quote must be checked.
- Input: live authorized brokerage facts
- Output: readiness assessment or blocked decision
- Boundary: every actual brokerage action requires explicit human confirmation
03 / Interactive decision tree
Every branch has a reason to advance or stop
Choose a decision gate to see the yes/no outcomes. This is an educational explanation of the architecture, not a live market evaluator.
Selected gate
Evidence quality
Validate issuer facts, event timing and market-data freshness before constructing a decision.
YES: Proceed to thesis and red-team evaluation.
NO: Research further, mark unknown, or reject the idea.
A confident conclusion cannot replace missing evidence.
04 / Controls and decision authority
Governance is part of the architecture
Controls vary by stage and rely on current facts. Illustrative rules from this private experiment should not be treated as universal investment recommendations.
Research gate
Every material idea needs a written thesis, entry criteria, invalidation, target, catalyst logic, and exit plan. Options must be compared against shares and cash.
Concentration gate
The framework limits exposure to 25% of verified account value per underlying across instruments, checked against current holdings.
Drawdown circuit breaker
A decline of 50% from a verified strategy high-water mark halts new or increased risk until a review. Historical records alone cannot verify current drawdown.
Human authorization
Research and analysis may be scheduled. Account-specific sizing and any brokerage execution require fresh verification and explicit user confirmation.
05 / Operating cadence
Observable autonomy, repeatable reviews
The documented operating model separates lightweight monitoring from more extensive research and lifecycle reviews. Operational health is monitored independently.
Public-market exception watch and workflow status heartbeat.
Unanchored opportunity discovery and candidate verification.
Lifecycle recap, maintenance observations and reconciliation.
Strategy reset, research quality review and process audit.
What persists between runs?
Research Pipeline, Investment Committee decisions, Portfolio lifecycle, Trade Log, Activity Log, Weekly Reviews and Account Metrics. Dates, source evidence, next review and exit-plan fields support continuity.
What happens when information is stale or missing?
Decision-critical uncertainty is recorded as a gap. Research may continue with suitable current public sources, but private account facts cannot be inferred from notes. The system escalates for an interactive review when required.
How are experimental ideas tested?
Shadow assessments can compare proposed logic without altering established decision history, risk gates or execution permissions. Experimental features are not represented as deployed production controls.
Why separate a health guard from research?
Scheduled-task health checks identify missing or degraded runs without expanding into market research, brokerage account access, or unrelated remediation.
06 / Public documentation boundary
Detailed by design. Sanitized by default.
Shared here
System responsibilities, data flows, source hierarchy, state transitions, decision gates, scheduled workflow patterns and reference governance concepts.
Kept private
Balances, positions, order history, real account identifiers, authentication secrets, private endpoints, internal task identifiers and live research records.
This is an independent educational engineering case study. It does not provide investment advice, real-time trading signals, automated trading access or a performance representation.
A Gericke Family project
Explore the broader collection
Technology services, creative work, and technical experiments, independent in purpose and connected by curiosity.
What We're Building →Technology Services →