CivicMind

Problem
Agentic AI Loop
Cover
Solution
Human-in-the-Loop
Phase 2 Monitoring Pipeline
System Architecture
Competitive Edge
Roadmap

Inspiration

Sri Lankan B2B teams often monitor competitors manually by checking company websites, news articles, career pages, product updates, pricing pages, and public announcements. This process is repetitive and time-consuming, and important updates are easy to miss.

A business may discover too late that a competitor has launched a new product, changed its pricing, started hiring for a new market, or announced an important partnership. Smaller companies may also be unable to justify the cost and complexity of enterprise competitive-intelligence platforms.

This inspired us to design CompetitorPulse LK, an affordable, evidence-focused competitive-intelligence agent for Sri Lankan B2B teams.

What it does

CompetitorPulse LK monitors selected public competitor sources and converts scattered market signals into structured, evidence-backed business intelligence.

The agent can:

Monitor competitor websites, news sources, pricing pages, and career pages
Compare current content with previously stored snapshots
Detect meaningful changes while filtering minor page noise
Classify changes such as product launches, pricing updates, hiring, partnerships, campaigns, or market expansion
Calculate confidence and business-importance scores
Explain why a detected change may matter
Recommend follow-up actions
Generate draft weekly competitive-intelligence briefs
Route alerts and tasks through human approval before execution

The system is not designed as a simple chatbot. It follows an agentic workflow in which it selects tools, gathers evidence, compares historical context, evaluates confidence, handles failures, and requests human review when information is uncertain.

How we designed it

For Phase 1, we designed a technically feasible architecture, agentic reasoning loop, Human-in-the-Loop policy, and one-week MVP implementation plan.

The planned workflow is:

Receive a competitor-monitoring goal
Read the configured competitor watchlist
Select the appropriate data-collection tool
Collect signals from websites, news, and career pages
Compare the collected data with stored baseline snapshots
Detect and filter meaningful changes
Classify the change type
Calculate confidence and importance scores
Check whether the available evidence is sufficient
Search for additional evidence when necessary
Generate an evidence-backed recommendation
Send the recommendation to a human review gate
Execute only approved alerts or tasks
Record the decision and execution result for traceability

The proposed architecture uses a React dashboard, FastAPI backend, LangGraph agent orchestrator, Playwright website collector, Tavily Search API, PostgreSQL with pgvector, Gemini reasoning, an evidence-aggregation layer, and approved Gmail or Trello actions.

Human-in-the-Loop

Human control is a central part of CompetitorPulse LK.

Before an important alert, report, or task is sent, a manager can:

Approve the recommendation
Edit the recommendation
Reject it
Request more evidence
Change the importance level
Reclassify the result
Mark it as a false positive

Low-confidence or conflicting information never triggers an external action automatically. All approved actions are recorded in an audit log.

Challenges we faced

One major challenge was ensuring that the system represents genuine agentic reasoning rather than a fixed automation pipeline.

We addressed this by including:

Dynamic tool selection
Confidence-based decisions
Evidence-sufficiency checks
Replanning when evidence is weak
Recovery from API and website failures
Duplicate-signal handling
Conflict detection between sources
Human approval before action execution

Another challenge was keeping the MVP realistic for a one-week build. Therefore, the initial scope is limited to 3–5 competitor targets and three signal types: websites, news/search results, and public career pages.

What we learned

Through this project, we learned that a useful AI agent must do more than summarize information.

It must:

Select the right tools for the current task
Compare historical and current data
Reason about business impact
Handle incomplete or conflicting information
Explain its conclusions using traceable evidence
Stop or replan when confidence is inadequate
Keep humans in control of high-impact actions

We also learned the importance of designing a modular and failure-tolerant architecture that can be adapted to the specific problem statement released during Phase 2.