Endeit Capital · Investment Analysis
Engineering notes ↗

AI-powered company verdicts with full source provenance.

Six specialist agents fan out from a typed company profile, the synthesis layer combines them into a recommendation, and a devil’s advocate stress-tests the bull case. Every cited fact traces back to its source URL — and when the data is too thin to judge, the system says so explicitly instead of recommending pass.

Pipeline

sources → typed profile ⇄ gap-filler → six specialists → synthesis → bear case
SourcesCrunchbaseLinkedInTavilyOwlerDealroomTrustpilotHiringProfileCited[T]Conflict[T] · None9 sectionsGap-filler · Firecrawlfinds gaps · enrichesenrichment notesSpecialistsTEAMARPROTRACOMFINSynthesisBull caseweighted scoreBearDA
Langfuse · live observabilityopen dashboard ↗
10 traced / 10 analyses
LLM calls
56
110 observations
Tokens
492.2k
prompt + completion
Avg latency
23.7s
per analysis
Total cost
free
Ollama Cloud
Models
gpt-oss:120b
in use

Concept philosophy

design principles that make the verdicts trustworthy
  • Provenance
    Cited[T]

    Every fact in the profile carries the URL it came from, when it was fetched, and the confidence we place in it. There are no orphan numbers.

  • Disagreement is data
    Conflict[T]

    When two sources disagree on a fact (headcount, founding year, valuation), we keep both as candidates. The agents reason over the conflict instead of silently picking one.

  • Refusing to judge
    Insufficient data

    If half or more specialist agents are low-confidence, the recommendation collapses to insufficient_data — distinct from pass. Saying 'we don't know' is honest.

  • Adversarial reasoning
    Devil's advocate

    After the bull case lands, a separate agent re-reads the same evidence and steelmans the bear case. The bear can downgrade — or upgrade — the final recommendation.

  • Adaptive ingestion
    Gap-filler agent

    Before the specialists run, an agent inspects the profile, names the biggest gaps, and picks 3-5 URLs to scrape. Each company gets a custom set of additional sources.

  • Deterministic where possible
    Weighted synthesis

    The overall score is a weighted average — fixed math, not LLM judgement. The model only writes the bull-case prose. Recommendation thresholds are explicit.

10 demo companies