The Problem
Synthesis at the Moment of Decision
Product managers in regulated, knowledge-heavy industries drown in context. They maintain sprawling Confluence pages, research repositories, regulatory trackers, and competitor dossiers — yet when a new product idea surfaces, they still make decisions from memory and gut feel.
The gap isn't access to information. It's synthesis at the moment of decision.
I saw this pattern repeatedly: a PM proposes a feature, it goes through three rounds of review, and only in the fourth meeting does someone surface the compliance regulation that kills it — a document that was already in the system. Spark2Build is my answer to that problem.
What I Built
A Knowledge-Grounded Intelligence Layer
PMs capture institutional knowledge into a Knowledge Hub. When they write a new product idea (a Spark), the AI cross-references it against that knowledge base in real time.
Key Product Decisions
Five Choices That Shaped the Product
1. Decomposed Confidence Scoring
PMs didn't trust a single "black box" AI score. I decomposed confidence into three transparent dimensions: Structural (entity co-occurrence in the knowledge graph), Semantic (vector embedding cosine similarity), and Temporal (decay based on entry age and domain-specific staleness thresholds).
2. Progressive Graph Modes
To balance speed and accuracy as workspaces grow, I built three modes: flat (pure semantic search, <30 entries), graph_lite (entity co-occurrence boost, 30–99 entries), and graphrag_full (full graph-aware retrieval, 100+ entries).
3. Regulations Force-Injection
Passive similarity search often misses implication-based links. Regardless of scores, I always inject the top 3 regulatory entries into context to catch compliance risks early — the primary reason to buy in financial services.
4. Proactive Radar Feed
A push-based feed that surfaces stale entries, semantic overlap when teammates add new knowledge, and whitespace "opportunity signals" — turning the platform from reactive to anticipatory.
5. Inline Contextual Surfacing
Real-time similarity search as the PM types context — surfacing relevant knowledge while intent is fresh, and capturing intent-signals for the confidence model.
Architecture
Choices That Reflected Product Thinking
Lessons
What I Learned About Enterprise Product
Compliance is a Feature
In financial services, surfacing a regulation early is the primary reason to buy the product — not a constraint to design around.
Explain the AI
PMs don't trust AI they can't explain. Decomposed confidence was the answer to "why should I believe this?" — transparency drives adoption.
Context Portability
Enterprise PMs often straddle multiple teams. Carrying knowledge isolation across workspaces was a huge differentiator.
Product Philosophy
"Insight without influence does not change decisions."
I prioritize narrative-driven data — ensuring every AI-generated signal is supported by enough evidence to influence a C-level decision.
"Spark2Build is what I'd want if I were still a PM in a regulated enterprise."
Always open to discussing AI strategy, enterprise scaling, or signal-to-roadmap conversion.
Get in Touch