A research agent that keeps 340+ notes findable — without a human curator.
Maintains a source-of-truth knowledge base with deep-read synthesis, hybrid search, and weekly self-healing. Every downstream marketing claim is validated against this KB — which makes fabrication structurally impossible.
AI-written content fails on one dimension more than any other: sourcing.
A marketing team publishes a confidently-worded post citing a statistic that was never published. A customer clicks the source link and finds a 404. A legal team asks where the number came from and nobody has an answer. The tool saved drafting time and spent it back on cleanup.
LLMs will invent citations. They'll quote figures from papers that don't exist, attribute claims to companies that never made them, and link to articles that have long since gone dead. By the time anyone catches it, the post is live and the damage is a brand-trust event, not a typo.
The root cause isn't the model. It's that there's no single source of truth the content gets validated against — no searchable, versioned, auditable set of files that downstream validators can compare a draft to. Sources live scattered across Slack threads, bookmarks, drive folders, and Notion docs. Redundant research is the default. Broken links accumulate.
This agent exists to turn that scatter into one searchable KB that grows with every session, heals itself weekly, and makes "where did this claim come from?" a one-query answer.
Routing-aware search. Deep reads, not snippets. Weekly self-healing.
The agent consults a routing map before it searches.
Before running any query, the agent reads meta/source-routing.md — a topic-to-file map built up over hundreds of sessions. "For AI agents, check these files first." This is what stops the 300-parallel-Google-search death spiral that kills most research automation.
Deep reads are mandatory. Snippet-skimming is forbidden.
The system prompt blocks shallow reads. For every question, the agent reads 5–10 full matched files, cross-references across products, tweets, and concepts, and only then synthesizes. Judgment lives in the cross-references, not in surface-level pattern matching.
Every session writes machine-readable files, not chat logs.
Insights go to wiki/insights/ with explicit citations. Query patterns, dead ends, and new topic mappings get logged to meta/. The next session reads this memory — so mistakes don't repeat and successful query patterns compound.
Sunday morning the KB heals itself.
A weekly cron runs KB Enhance: broken links get repaired, missing concepts auto-generate, contradictions get flagged for human review. The KB improves without human maintenance — which is the only way a 340-file knowledge base stays trustworthy over time.
Three things change once the KB is live.
Indexed and queryable KB
Mix of product docs, curated tweets, foundational articles, and 500+ word syntheses — all machine-searchable via hybrid (vector + keyword + BM25) retrieval.
Marketing claims validated against the KB
Claim-coverage and source-freshness gates depend entirely on files the research agent produced. Hallucinated stats have no way to pass.
Automated KB maintenance jobs
KB Enhance, lint, compile, and concept auto-generation run on a schedule. Broken links get fixed without a human remembering to look.
Numbers observed in Brilworks' internal reference deployment. Actual figures on your stack will depend on source volume, publishing cadence, and how strict you want the validator gates.
Honest fit criteria. We'd rather say no than oversell.
✓Strong fit if
- You publish AI-written content and can't afford fabricated citations reaching the public
- You have curated sources scattered across Slack, drive, bookmarks, and Notion and no single search surface
- You want a KB that improves with every session, not a static repository
- You need an audit trail showing exactly where each published claim came from
✗Not a fit if
- You publish one post a week (the KB maintenance overhead is wasted)
- You have no curated sources yet — start with positioning and brand, not a KB
- You want fully autopublished content with no validator gates in front of it
- All your sources are public web URLs with no proprietary curation
Book a 30-minute scoping call.
We'll walk through your current sources, map them into a searchable knowledge base, and show you how to wire validators on top — then tell you honestly whether it fits.
+ 91 9313644148
sales@brilworks.com
503, Fortune Business Hub, Science City Road, Sola, Ahmedabad, Gujarat, India Pincode- 380060
