TradeSense AI: Agentic Investment Assumption Monitoring Platform

Executive Overview

Private-market investment firms (venture capital and growth equity) invest billions based on detailed investment memoranda. Once a deal closes, these memoranda become static documents while the world changes dynamically around them: competitors emerge, regulatory environments shift, customer unit economics deteriorate, and core market assumptions weaken.

TradeSense AI was architected to solve this disconnect by transforming static investment theses into live, continuously monitored Assumption Graphs that track evidence, detect emerging contradictions, and generate materiality alerts for investment committees.

The Architectural Challenge

Most "agentic" prototypes rely on naive iterative loops that fail under enterprise requirements:

1. Unbounded Autonomy Risk: An autonomous agent that can unilaterally decide a portfolio company's thesis has failed poses unacceptable governance risk.
2. Fragile Long-Running Workflows: Deep evidence research across SEC filings, patent databases, earnings calls, and news takes minutes to hours. If a server restarts or hits a timeout, traditional stateless LLM chains lose all execution state.
3. Hallucinated Evidence: General-purpose LLMs synthesize plausibility rather than verifiable citations.

The Solution & System Design

1. Specialized Multi-Agent Orchestration

Rather than relying on a monolithic prompt, TradeSense AI employs a specialized Orchestrator/Planner coordinating modular worker agents built on LangGraph subgraphs:

• Extraction Agent: Deconstructs uploaded PDFs into atomic, testable assumption nodes.
• Evidence Research Agent: Queries financial APIs, news feeds, and vector indices with autonomous retrieval depth planning.
• Contradiction / Red-Teaming Agent: Actively searches for disconfirming evidence against the original thesis.
• Materiality Reasoning Agent: Evaluates whether discovered evidence fundamentally shifts the investment thesis risk profile.
• Critic & Verifier Agent: Enforces strict citation grounding before any output is synthesized.

[ Investment Memo / PDF ] 
            │
            ▼
   ┌──────────────────┐
   │ Extraction Agent │ ──► [ Atomic Assumption Graph ]
   └──────────────────┘                   │
                                          ▼
   ┌─────────────────────────────────────────────────────────────┐
   │              Orchestrator & Research Engine                 │
   │  ┌────────────────────┐          ┌───────────────────────┐  │
   │  │  Evidence Search   │ ◄──────► │ Contradiction Red-Team│  │
   │  └────────────────────┘          └───────────────────────┘  │
   │                             │                               │
   │                             ▼                               │
   │                ┌─────────────────────────┐                  │
   │                │  Materiality Evaluator  │                  │
   │                └─────────────────────────┘                  │
   └──────────────────────────────┬──────────────────────────────┘
                                  │
                                  ▼
                    ┌───────────────────────────┐
                    │ Structural Human Boundary │ (propose_status_change)
                    └─────────────┬─────────────┘
                                  │
                         [ Investment Partner ]
                                  │ (Approve / Reject)
                                  ▼
                    [ Live Portfolio Dashboard ]

2. Structural Human-in-the-Loop Governance

A core architectural innovation in TradeSense AI is structural rather than prompted governance. The agent tool registry exposes propose_status_change but deliberately omits commit_status_change.

Any transition toward a Weakening, Review Required, or Invalidated thesis status is architecturally forced through human review. A hallucinating agent cannot mutate portfolio status by design.

3. Durable Execution with Temporal.io

To support long-running research cycles that pause for human approval, the orchestration tier uses Temporal.io. Workflows can suspend execution indefinitely while waiting for an investment partner's decision, consuming zero compute resources, and resume deterministically with zero state loss.

Technical Stack & Performance

• Agent Tier: Python, FastAPI, LangGraph, Pydantic v2
• Platform Core: NestJS, TypeScript, PostgreSQL with pgvector
• Workflow Engine: Temporal.io
• Event Streaming: Apache Kafka for asynchronous document ingestion
• Deployment: Docker, Kubernetes (EKS), Terraform

Results

TradeSense AI delivers evidence-linked briefs in under 60 seconds, reduces manual portfolio monitoring overhead by 70%, and provides investment partners with verifiable, audit-proof oversight across their portfolio companies.