2026 – present

echosystemai

what it is

echosystem turns news about a company or topic into a knowledge graph and a short written report, then checks that report against the articles it came from.

collect
extract
connect
draft
audit

my favorite part

when two articles disagree, the graph doesn't pick a winner. it keeps both claims and links them as a contradiction, so the conflict stays visible.

what i built

this one is a solo project, from the first commit to the dashboard.

ingestion: a background worker pulls up to 8 recent articles per topic, cleans them and splits them into chunks.
extraction: a small language model finds entities and relationships, and every result is validated against a fixed schema.
the graph: entities, relationships and claims go into neo4j, and each new claim gets checked against the existing ones for contradictions.
the audit loop: a langgraph workflow where a researcher drafts and a separate auditor returns a structured verdict.
the app: a fastapi backend and a next.js dashboard for reports, a watchlist and open questions.

see it running

the echosystem dashboard focused on games workshop, with lines to amazon, united artists, mike flanagan and henry cavill, and a panel listing each relationship with its source sentence — click a company and the graph shows who it’s tied to, with the sentence behind each link. this is real data from an article about warhammer adaptations. the contradiction badges come from the rough keyword check, so don’t read much into them yet.

two reports in the dashboard: a boeing 737 report marked pass and an nvidia report marked fail — two real reports. the boeing one passed on its second audit, with claims tied to its sources. the nvidia one wrongly said acer acquired nvidia, failed all three audits and was labeled unverified. a pass still only means the auditor found nothing wrong. captured from a local run in november 2026.

why a graph and vectors

vector search is good at finding relevant passages, and the graph makes the connections between companies explicit and easy to follow. so retrieval takes the top 5 vector hits, then pulls up to 10 graph neighbors for each hit's main entity.

when the auditor says no

the auditor returns pass or fail, a reason, and the claims it flagged. on a fail, the researcher redrafts with that reason added to its prompt. after three failed audits, the report is saved with an unverified note instead of being passed off as checked.

real output, netflix run

There is no information in the context about Netflix's current market position or future plans.

[UNVERIFIED: failed audit after 3 attempts. Review manually.]

the articles it pulled were mostly about warhammer adaptations, not netflix. all three drafts failed the audit, and the report says so.

the honest bit

tested: automated tests cover the audit routing, the contradiction logic and the report storage.
not measured yet: i don't have labeled data, so how often the auditor or the contradiction check is right hasn't been measured. a pass means the model found nothing wrong, not that the report is true.
known limits: contradiction detection is keyword-based by default, and entities match by exact name, so “nvidia” and “nvidia corp” become two nodes.

built with

python
langgraph
neo4j
qdrant
fastapi
celery
redis
next.js

see the code on github ← previous project next project →