Overview
Venture capital runs on intuition disguised as data.
I built TractionX to give that intuition structure — an agentic OS that analyzes founders, startups, and markets through the lens of investor theses.
The idea was simple: instead of forcing investors to read every deck,
let the system learn what they care about and surface what actually matters.
But the execution meant building a new kind of data and AI stack from the ground up.
The Challenge
Investors were drowning in noise: decks, forms, CRMs, spreadsheets, and half-clean data from enrichment APIs. Everyone had access to data - few had signal.
We needed to:
Aggregate startup data from multiple sources (forms, pitch decks, APIs).
Map that data to each investor’s thesis - geography, sector, stage, business model.
Score alignment automatically and explain why.
Generate reports that an IC could actually use.
And it had to do this reliably, transparently, and fast - not as a demo, but as a real workflow.
The Solution
TractionX became a modular OS with three core layers:
Data Layer - Async pipelines pulling founder, company, and signal data from sources like Apollo, Crunchbase, Clay, and custom scrapers. All normalized into structured PostgreSQL models and enriched with embeddings in Qdrant.
Thesis Layer - Every investor defines matching and scoring rules per question or attribute. The system interprets those rules through a mix of symbolic logic and LLM reasoning. It doesn’t just say “match 80%” - it explains why.
Intelligence Layer - Generates proprietary IC reports using natural-language reasoning, backed by fully traceable data. Redis queues manage concurrency; LangFuse traces every token and job.
The UI exposes results like an AI console: scores, rationales, and insights - all rendered in seconds.
The Impact
$100K / 6-month pilot contract with a Singapore-based internal VC.
3 paid investor pilots and 15+ beta orgs onboarded across SEA and Europe.
Reduced diligence cycle time from weeks → hours.
Generated consistent, audit-ready scoring across hundreds of startup submissions.
It proved that the pain point wasn’t lack of AI - it was lack of structure. Once you model structure, AI becomes inevitable.
The Learning
I learned that “AI for investing” isn’t about predictions, it’s about context. Data without context is noise; context without structure is opinion. Building TractionX taught me that clarity scales better than complexity.
It’s not a tool, it’s a system. And systems, when designed well, become extensions of thought.
Tech Stack
Python · FastAPI · Redis (RQ) · PostgreSQL · S3 · LangGraph · OpenAI · Qdrant · LangFuse · AsyncIO
Status
Private Beta — Active Development
Pilot deployments live with early investors in SEA and Europe.