Case study · 2026
Opportunity Radar
A research pipeline that turns scattered public signals into traceable, budget-aware opportunity analysis.
- Role
- Product designer, architect, and operator
- Outcome
- A resumable multi-stage pipeline with versioned configuration, model tiering, scheduled scans, and explicit API cost gates.
- Capabilities
- Applied AI · Product engineering · Pipeline reliability
- Stack
- Go · SvelteKit · PostgreSQL · Redis · Ollama
Interactive system map
One scan, six durable boundaries
Illustrative flow · no external requests
Why this boundary exists
Expand
Turn a research prompt into focused concepts and query families.
Case record
The problem
Useful opportunities rarely arrive as clean records. They emerge across discussions, releases, complaints, and technical shifts. The tool needed to gather those weak signals without presenting model output as unexplained truth.
The constraints
- Search and scraping fail independently and unpredictably.
- Local models and paid models have different strengths and costs.
- A long scan should resume from its last durable phase.
- Prompt and threshold changes must remain attributable to a specific run.
Pivotal decisions
The pipeline separates expansion, search, scraping, classification, extraction, and synthesis. Each stage persists its result and publishes progress, so retrying a failed run does not repeat completed work.
Fast local models filter and structure the raw material before a stronger paid model synthesizes conclusions. Daily and monthly budget gates stop the final stage before it can incur unplanned spend. Immutable configuration versions preserve the prompts, thresholds, and query templates used by every scan.
What it demonstrates
Applied AI is treated as an operated system: bounded, observable, resumable, and designed around uncertainty rather than a single successful prompt.