The intelligence stack
Built around a single insight:
developers discover what the market wants next.
Every layer of TrendIntel is purpose-built to surface emerging opportunities at Stage 0–2 — when the signal exists, the market does not, and the window is at its widest.
Data Ingestion
809 sources. One intelligence layer.
TrendIntel ingests signals from across the internet — developer communities, academic research, startup activity, job markets, social platforms, and mainstream media. Every source is weighted by community type to reflect true signal quality, not volume.
Signal Weighting
Not all signals are equal. The weights reflect that.
A developer discussing a problem in a technical community is a stronger signal than a journalist writing about it. Community weights encode this hierarchy into every score.
Academic
3.0×Research preprints are the earliest signal — problems are identified before solutions exist.
Developer
2.5×Developer adoption precedes startup formation by months. Where builders look, markets follow.
Startup
1.8×Product launches and startup formations confirm a market is emerging from developer interest.
Consumer
1.5×Consumer discussion signals broader adoption — important context, not a leading indicator.
Institutional
1.5×Government contracts, legislation, and regulatory filings. Real capital committed is a structural adoption signal.
Market Demand
1.2×Job postings and hiring signals. Companies actively recruiting for a technology confirms production adoption.
Mainstream Media
1.0×Media coverage is a lagging indicator. High weight here means the opportunity is closing.
Propagation Stages
Know exactly where a trend is in its lifecycle.
Six propagation stages map the journey from underground signal to mainstream saturation. The stage tells you how much whitespace remains and how much time you have.
Academic and isolated developer signals are appearing. The problem is identified but no product or community momentum exists yet. The highest-potential and hardest-to-act-on stage — requires conviction.
Developer communities are actively discussing and building. Code repositories, Q&A questions, and package adoptions are spiking. This is the primary target stage — whitespace is high, signal is real.
Product launches, seed funding, and technical job postings are appearing. The market is forming. Still time to build — but the clock is running.
Forum threads, tutorial content, and community discussion are proliferating. Consumer awareness is building rapidly. Execution speed is becoming the primary differentiator.
Search volume spikes. The general public is searching. Multiple funded competitors exist. Entry is still possible with a strong differentiation thesis, but the opportunity score has peaked.
Mass media coverage. Consumer purchase velocity spikes. The trend has completed its propagation cycle. Opportunity scoring drops significantly — saturation is high, whitespace is minimal.
Scoring Engine
Six-dimensional opportunity scoring, updated hourly.
Each topic is measured across six orthogonal dimensions running in a chained pipeline every hour. Together they capture both current state and forward trajectory.
Velocity
(current_week / avg_weeks_2–7) − 1
Community-weighted signal count ratio over 8 rolling weeks. Acceleration is measured relative to the baseline period so short spikes are distinguished from sustained growth. Academic and developer signals carry higher multipliers.
Stage Detection
Signal composition thresholds → Stage 0–5
Assigns a propagation stage based on community type composition. Developer and academic signals dominate early stages. Consumer and media signals dominate late stages. Stage determines whitespace and timing context for every score.
Saturation
product(30%) + startup(25%) + media(20%) + age(15%) + diversity(10%)
Measures how crowded the space already is. High saturation means the market is mature. Low saturation with high velocity is the target combination — it quantifies whitespace as a single actionable number.
Opportunity
demand(30%) + problem_density(25%) + velocity(20%) + whitespace(15%) + stage(10%)
The headline score for builders and decision-makers. Combines demand signals, detected user complaints (problem density), acceleration, and market whitespace. Triggers AI brief generation when it crosses the minimum threshold.
Momentum
velocity(35%) + acceleration(25%) + diversity(20%) + cross-platform(10%) + formation(10%)
Measures the quality and breadth of growth. Cross-platform spread and product formation signals are leading indicators. A trend accelerating across multiple source types scores much higher than one confined to a single community.
Predictive
momentum(30%) + trajectory(25%) + whitespace(20%) + stage_velocity(15%) + opportunity(10%)
Forward-looking explosion probability. Answers: will this trend be everywhere in 6 weeks? Requires a minimum signal history to activate. Incorporates external forecasting consensus where available.
Semantic Clustering
Topics emerge from data, not from human categories.
Every signal is embedded into a high-dimensional semantic space using a multilingual model trained on 50+ languages. A forum post in English and a research abstract in German about the same problem cluster together correctly — no translation required.
Density-based clustering groups signals without requiring a predetermined topic count. No one decides in advance what categories exist — topics surface from the signal distribution itself. This is how genuinely new categories become visible before they have names.
A narrative detection layer runs weekly, identifying higher-order meta-trends that span multiple individual topics and represent structural shifts rather than isolated spikes.
Signal processing pipeline
Cluster continuity
New clusters are merged with existing topics when their centroids are sufficiently close in semantic space — preserving 90-day trend history across clustering runs.
Reclustering triggers automatically when unassigned signal volume crosses a threshold, keeping topic boundaries fresh without manual intervention.
Micro Trends
Earlier than Stage 0. The embryonic layer.
The main clustering pipeline groups signals into topics once enough semantic mass accumulates. The signals that don’t fit — the ones the density algorithm classifies as noise — are not discarded. A second clustering pass runs against that noise, finding embryonic signal clusters too new and too small to surface in the main pipeline. By the time a topic earns Stage 0, a micro trend has often been visible for weeks.
Each micro trend is tracked across clustering runs by centroid proximity, so you can see whether it is holding together, dissolving, or growing. Velocity is measured against its own history — a micro trend accelerating from 6 to 18 signals in two weeks is a fundamentally different signal than one holding flat at 8.
When a micro trend’s centroid converges on an existing topic cluster, it is automatically promoted into the main pipeline, triggering full scoring, AI brief generation, and all standard tracking. Watchers are notified at the moment of promotion so they can act before the broader feed surfaces it.
Illustrative data — live dashboard shows real-time micro-clusters
AI Analysis
Analyst-level opportunity briefs. Automated nightly.
When a topic crosses the opportunity score threshold, TrendIntel dispatches an AI analysis job that generates a structured brief. No manual trigger. No queue. Every qualifying topic, every night.
Each brief identifies the core unmet need, generates 3–5 specific product hypotheses, surfaces the primary market risks for new entrants, and ranks the top monetization models by fit for this specific opportunity type.
Our custom AI analysis engine is purpose-built for market intelligence — not a general-purpose summarizer. The output structure is consistent, comparable, and designed for high-stakes decision-making.
Developers using AI for code generation lack structured tooling to version-control and semantically review AI-generated changes at PR time. Standard diff tools treat AI output like human output.
- ›Semantic diff layer for AI-generated PRs with intent classification
- ›AI change attribution system integrated into version history
- ›Code review assistant trained on AI-generated code patterns
AI Trend Chat
Ask pointed questions. Get specific answers.
Every trend page includes a conversational AI interface with full context: the topic's current stage, velocity score, opportunity score, signal composition, and the AI-generated brief. Ask it anything specific to that trend.
This is not a general-purpose chatbot. The AI knows it is talking about a specific emerging market, at a specific stage, with a specific opportunity profile. The answers reflect that context.
Use it to pressure-test an opportunity, explore adjacent markets, stress-test a startup idea against the signal data, or understand what is actually driving a velocity spike.
Example narratives — live data varies
Macro Narratives
Structural shifts, not just topic spikes.
Individual topic clusters tell you that a specific technology is gaining traction. Macro narratives tell you something bigger: that multiple clusters are converging on the same structural shift at once.
Narrative detection runs weekly across all topic embeddings, identifying groups of topics that are semantically related and co-accelerating. Each narrative is scored by strength — the number of constituent topics, their combined velocity, and the coherence of the underlying signal.
A high-strength narrative is the earliest signal of a category-defining market forming — the kind of shift that produces multiple large companies, not just one.
Signal Search
Read the raw evidence behind every score.
Every score in TrendIntel is derived from real content: actual posts, questions, job listings, and research abstracts that were ingested, embedded, and assigned to a topic. Signal search gives you direct access to that underlying evidence.
Search and filter by keyword, source type, signal type, or topic assignment. Read the original text. Understand what specifically is driving a velocity spike, a problem density signal, or a stage transition.
When you are deciding whether to act on an opportunity, the question is not just what the score says — it is whether the underlying signals hold up to scrutiny. Signal search is where you do that work.
Has anyone found a good way to review AI-generated PRs at scale? Our review process is breaking down when 60% of the code is AI-authored.
Semantic Change Attribution in Large Codebases: A Framework for AI-Assisted Development Workflows
The fundamental issue is that git diff has no concept of intent. Two AI-generated changes that look identical syntactically can be semantically opposite.
Entity Intelligence
Track the names the market is starting to say.
Every signal TrendIntel ingests is run through named entity recognition, extracting the specific companies, products, people, places, and events embedded in the text. These entities are then tracked for velocity, source diversity, and topic cluster spread — in real time.
The result is a second lens on the signal stack: not just what topics are emerging, but which names are suddenly appearing across sources and conversations that have never been connected before. A company showing up simultaneously in arXiv papers, Reddit threads, government contracts, and job listings is telling you something before any analyst has written a word about it.
Entity types tracked
Rising Entities — sorted by velocity
Entity Communities
When two stable entities start moving together, the collab is already forming.
Single-entity velocity misses the most interesting trend pattern of all: two or more well-known entities suddenly being talked about together in ways they never were before. Neither one's individual mention count needs to spike. The signal lives in the co-occurrence.
TrendIntel computes pair-level co-occurrence velocity hourly across every signal we ingest, then runs Louvain community detection over the resulting weighted graph nightly. The output is a feed of emerging entity communities — tight clusters of 3–10 entities that are converging into a single conversation: brand collaborations, product ecosystems, competitive landscapes, or parallel narratives.
Each community is AI-labelled with the theme that connects its members, scored by an emergence metric that combines velocity, member count, and signal volume, and rendered as an interactive force-directed graph showing how the entities link together — with one-hop context neighbours so you can see where the conversation is happening.
4 entities · 21 signals/7d
12-week co-mention volume — synchronised lift-off in the last 3 weeks
Community graph
Two or more brands suddenly co-mentioned in product, lifestyle, or fashion signals. Surfaces drops and partnership announcements during the leak phase.
Multiple vendors in the same category clustered together — e.g. cybersecurity firms co-discussed in vendor comparisons, M&A speculation, or category-defining content.
A product line, a tournament roster, a music lineup, a regulatory bundle — any set of entities being discussed as one conversation, surfaced by the graph.
Institutional Intelligence
Social tells you what's moving. Filings tell you who's betting on it.
Every other layer on this page reads the grassroots — forums, code, research, social. Institutional Intelligence reads the other side of the market: what public companies formally tell regulators. We ingest long-form SEC filings, parse their narrative sections, and run the same NLP stack over them that powers the social pipeline.
Coverage spans the full institutional record. On the narrative side — 10-Q and 10-K reports, 8-K material events, S-1 and 424B offerings, 20-F foreign annual reports, DEF 14A proxy statements, earnings-call transcripts, Federal Reserve (FOMC) statements and minutes, and litigation dockets — each document is mined for named entities, AI-grouped into themes, and scored sentence-by-sentence by a finance-tuned AI sentiment model, so you see not just who is named but how favorably they're framed. The capital-flow filings — 13F institutional holdings, Form 4 insider transactions, and 13D/13G activist stakes — add the positions and insider moves behind the words.
The payoff is cross-source conviction: every trend is scored on how many independent worlds back it at once. A trend with grassroots momentum and institutional money behind it is a fundamentally different bet than one trending on social alone — and the panel shows you the exact companies that bridge the two, clickable straight through to the evidence.
Theme · 14 filings · last 30 days
Entity heat · framing
Cross-source conviction
Edge AI Inference
Entities clustered in filings around supply constraints, litigation, regulation, or margin pressure — what companies are collectively warning about, before it hits the headlines.
Entities discussed in the context of new markets, demand inflection, or product expansion — where institutional framing is turning positive across multiple filers at once.
Capital and capacity commitments — capex, partnerships, acquisitions — naming the same set of companies, mapped straight to the social trends they corroborate.
Exclusive signal layers
Signals no competitor monitors
Most trend tools watch social and search. TrendIntel adds the upstream layers that predict what those channels will say months from now.
Government procurement
Federal contract awards (USASpending.gov), SBIR/STTR grants, NIH and NSF research funding, Federal Register proposed rules, and Congressional bills — the full layer of government activity that precedes commercial markets.
No other trend platform ingests this data layer.
Prediction market signals
Metaculus, Manifold Markets, and Polymarket aggregate probability estimates from domain experts on specific future outcomes. Rising conviction on a topic in prediction markets often precedes narrative formation in mainstream media by weeks or months.
Unique to TrendIntel in this category.
Academic & patent signals
arXiv preprints, Semantic Scholar citation velocity, Hugging Face model and dataset trends, Google Patents via SerpAPI, and USPTO trademark filings — the research layer that precedes everything else by 1–3 years.
Weighted at 3.0× in the scoring pipeline.
The pipeline is running right now.
Get started free and see what is emerging today.