We built this for ourselves first.
That's the honest version of how Xenith Intelligence Suite got started. We were running Xenqube across blockchain engineering, AI product development, and cybersecurity — and every Monday started the same way: a 2-3 hour research session trying to answer two questions. Where is the market actually moving this week? And do we have the credentials to win there?
Manual. Time-consuming. Opinion-driven. Stale by Wednesday.
So we built a system to answer both questions automatically, with live data, scored rigorously, and cited so we could defend the decisions to clients and investors. We called it Xenith Intelligence Suite — a Revenue Intelligence Platform. Four interconnected modules. One shared Postgres database. One weekly pipeline that runs automatically and tells us where to hunt.
This is a technical breakdown of how it works, what we got wrong building it, and what it teaches you about AI-assisted market intelligence for B2B services firms.
The Problem It Solves (Precisely)
If you run a services firm in blockchain or AI, you've felt this:
You have 12 potential market domains you could pursue. Some of them sound hot — everyone's talking about AI agents, RWA tokenization, stablecoins. But "everyone's talking about it" is not the same as "there's actual demand you can capture." The vendor noise in this space is enormous.
You also have your own credentials — years of project delivery, a website with case studies, conference talks, internal documentation. But how much of that actually maps to what the market is currently paying for? Can you credibly win a stablecoin infrastructure contract? Or would you be winging it?
Manual research fails here because it's slow (by the time you finish, signals have changed), subjective (whose opinion on which markets are hot?), and disconnected from your own capabilities (great market signal means nothing if you can't deliver).
Xenith Intelligence Suite answers two questions with data:
- Where is demand actually growing? (Xenith Radar)
- Can you credibly win there, and what should you build if not? (Xenith Verdict)
With two supporting modules:
- Xenith Capture — turns your conference talks, demos, and webinars into searchable capability evidence
- Xenith Pursuit — discovers and scores actual prospects from government tenders, Web3 funding, and hiring signals
Module 1: Xenith Radar
What It Does
It tracks 12 market domains (7 blockchain, 5 AI) using 5 independent live signals per domain:
- Hiring demand — job count from Adzuna API + RemoteOK keyword matching
- Capital flow — TVL (blockchain via DefiLlama) or market cap (AI coins via CoinGecko)
- Developer momentum — GitHub repository counts for domain keywords, filtered to last 90 days
- Funding signal — TechCrunch + crypto RSS mention count combined with disclosed funding amounts
- Public interest — Google Trends 12-month mean across up to 3 keywords per domain
Each signal is min-max normalized across all domains (so scores are relative, not absolute). They're combined into a composite score weighted by default:
hiring: 30%
capital: 25%
developer: 20%
funding: 15%
public: 10%
Missing signals don't penalize — they're excluded from the weight sum and the denominator adjusts accordingly.
Why This Works Better Than Opinion
Before this system, our market positioning was driven by what felt hot, what clients were asking for, and what we'd read recently. The problem: blockchain markets move fast, and what felt hot 6 weeks ago might have already peaked. AI agents might be getting funding but not hiring yet — or vice versa.
Hiring demand is a leading indicator of real purchasing intent. Companies don't hire for markets they aren't spending in. Developer momentum shows where the open-source activity is concentrating. Capital flow shows where money is going. Combining them reduces false positives significantly.
The Narrative Layer
After the algorithmic scoring, we run an LLM pass over the top 3 domains to generate executive summaries for leadership. This is where we had to be careful.
LLMs are confident even when wrong, and the executive summary is what gets shared in board prep. So we built validation rules: any number the LLM mentions that's greater than 5 must exist verbatim in the input data. Top 3 narrative domains must match the top 3 by composite score. One automatic retry on validation failure.
This eliminates the most dangerous failure mode: a confident, plausible-sounding summary that cites numbers the model invented.
Module 2: Xenith Verdict
This is the part that makes Market Depth actionable.
Knowing that "LLM Application Engineering" scored 78/100 means nothing if you can't deliver in that space. Build Readiness cross-references market signals with your own company's proven capabilities.
How It Indexes Company Depth
It starts with a same-origin BFS crawler that walks your website (up to 120 pages). All text is chunked, embedded with text-embedding-3-small, and stored in Postgres as JSONB vectors.
Simultaneously, it runs deterministic keyword extraction across all pages against an 8-category capability taxonomy (~85 keywords covering blockchain core, DeFi, stablecoins, RWA, AI/ML, AI governance, gaming/metaverse, and enterprise infrastructure). Each capability gets a depth score:
depth_score = min(100, evidence_count × 8 + len(sample_urls) × 5)
So a capability mentioned on 5 pages with 3 sample URLs scores much higher than one mentioned once.
How It Computes Verdicts
For each market domain, it does two things:
Capability match scoring — RAG retrieval of the top 8 most relevant chunks, scored by cosine similarity and keyword overlap with the domain query. Combined with keyword taxonomy matches:
cap_score = min(100,
keyword_score × 0.35
+ semantic_avg × 100 × 0.45
+ chunk_hits × 8 × 0.20
)
Opportunity scoring — blends market demand with company capability:
opportunity_score = market_score × 0.55 + cap_score × 0.45
Algorithmic verdict thresholds:
| Verdict | Condition |
|---|---|
| GO | opportunity ≥ 65 AND cap_score ≥ 45 |
| CONDITIONAL | opportunity ≥ 45 AND cap_score ≥ 30 |
| NO_GO | otherwise |
The LLM then writes a grounded verdict using only the retrieved evidence — including specific build plans (phases 1-3) and a revenue path. The LLM verdict can override the algorithmic label, but only based on cited evidence from RAG.
Why We Separated GO and CONDITIONAL
CONDITIONAL is the most useful verdict. It means: this market is hot and you have some depth, but there's a gap between where you are and where you need to be to win consistently. CONDITIONAL verdicts become a backlog for deliberate investment — new service pages to write, case studies to create, hires to make, talks to give.
GO verdicts go to the sales pipeline. CONDITIONAL verdicts go to the capability-building roadmap. NO_GO verdicts save you from embarrassing pursuits in markets where you'd be winging it.
Module 3: Xenith Capture
This module exists because we kept running into a recurring problem: company websites underrepresent actual capability.
Conference talks, client demos, and webinar recordings contain deep technical evidence that never makes it into website copy. But those videos sit on YouTube or behind JS-heavy marketing pages, inaccessible to a web crawler.
Xenith Capture fixes this.
It handles:
- YouTube, Vimeo, TikTok, Twitter/X with yt-dlp
- Generic marketing pages with Playwright (Chromium, 90-second timeout, 6-viewport scroll, network request capture)
- Direct video file URLs
For each video, it extracts audio (max 10 minutes), transcribes with local faster-whisper (small model, CPU, no API cost), and runs an LLM summarization pass to produce: one-line summary, key points, chapters with timestamps, notable quotes, and topics.
Completed transcripts are then indexed into Build Readiness RAG as source_type = "video" — enriching capability evidence for domains where the website coverage is thin. If you gave a talk about stablecoin architecture at a conference last year but never wrote it up as a case study, that talk now becomes evidence in your capability profile.
Module 4: Xenith Pursuit
Market depth tells you where to go. Build Readiness tells you if you can win there. Lead Intelligence finds the actual humans and organizations to pursue.
It pulls from 11 sources:
- Government tenders (SAM.gov, Grants.gov, World Bank, TED EU, GeM India, CPPP India)
- Web3 funding (DefiLlama raises, Web3.career hiring signals)
- AI launches (Product Hunt, YC AI directory)
- Startup funding (Crunchbase RSS)
Each lead gets a 7-component composite score: market demand fit, capability fit, keyword fit, budget indicator, recency, urgency (deadline proximity for tenders), and contact confidence.
The intelligence bridge reads Market Depth and Build Readiness verdicts to classify each lead:
- Demand-oriented — hot market, proven depth (GO or CONDITIONAL + capability ≥ 35): pursue now
- Future-oriented — hot market, current capability gap: track, build depth, engage when ready
- Generic — weak market or capability signal: deprioritize
Outreach drafts are LLM-generated, but human-reviewed before sending. No auto-spam. This matters for brand.
What We Got Wrong Building This
First version had no normalization.
We computed raw scores per domain and compared them directly. The problem: hiring demand in "LLM Application Engineering" is measured in thousands of job postings; public interest in "RWA Tokenization" peaks at maybe 30 on the Google Trends index. Direct comparison was meaningless. Min-max normalization across the domain set fixed this, but it also introduced a relative problem — scores change when you add or remove domains. Something to be aware of.
We underestimated website crawl quality.
The first RAG index we ran on a client's website was almost useless — the crawler got through 50 pages and found mostly navigation text, legal disclaimers, and SEO filler. Real capability evidence was buried in case studies, technical blogs, and service description pages. We added minimum text length filtering (80 characters per page) and tuned the BFS traversal to prioritize paths that looked like content rather than infrastructure. Crawl quality is everything for the capability scoring to work.
LLM verdicts need strict grounding or they hallucinate confidence.
In early testing, the LLM would write confident verdict statements citing specific project counts or team sizes that weren't in the input data. We added explicit validation: any number > 5 in the output must exist in the input. One retry on failure. Fallback to algorithmic verdict only if the retry also fails. Grounding is not optional when the output feeds executive decision-making.
The Weekly Pipeline (How We Actually Use It)
Monday morning, before the week starts:
- Xenith Radar runs a full refresh — all 5 signals, all 12 domains
- Build Readiness re-indexes if the website changed (new service pages, case studies)
- Lead Intelligence discovers and re-scores the lead pipeline
By 9am Monday, we have: a scored market report with top 3 domain narratives, updated GO/CONDITIONAL/NO_GO verdicts, and a ranked lead queue with demand-oriented and future-oriented buckets.
BD review focuses on demand-oriented leads first. Strategy conversation covers CONDITIONAL verdicts — which ones to invest in this quarter. Future-oriented leads inform the hiring roadmap.
This took us from "3 hours of Monday morning research" to "20-minute review of structured intelligence."
What This Means for B2B Tech Services Firms
If you're running a services business in blockchain or AI, the pattern we're describing applies directly to you — regardless of whether you use this specific system.
The foundational problem is the same: you're chasing a moving market with capabilities that are partially documented and mostly scattered across websites, documents, and videos that no one has synthesized. You're making positioning decisions based on gut feel and vendor noise.
The solution direction is also the same: live market signals (not analyst reports from 6 months ago), deterministic capability indexing (not a partner's opinion), and cross-referencing the two to get verdicts that actually support decisions.
We built Xenith Intelligence Suite because we needed it. We're now offering it to other firms who have the same problem.
What We Can Build for You
If you're looking at this and thinking "we need something like this" — there are two ways we can help:
We can deploy Xenith Intelligence Suite directly for your firm. Configure it for your market domains (we support any combination of blockchain and AI verticals, and can extend to others), index your website and video archive, and set up the weekly pipeline. You get ongoing GO/CONDITIONAL/NO_GO verdicts aligned to your actual market positioning.
We can build custom intelligence infrastructure for your specific needs — whether that's a market monitoring system for a fund, a capability audit tool for an enterprise, or a lead discovery pipeline tuned to your sector.
Either way, the core work is the same: turning market data and your company's knowledge into defensible, regularly-refreshed, decision-ready intelligence.
Xenith Intelligence Suite is currently in working POC (v1.0). We're deploying it with select services firms in the blockchain and AI verticals. If you're running a technical services firm and want to see how this maps to your positioning challenges, let's talk.