LeadLyft
Months of user activity turned into reports that read written-by-coach, not by-model — three AI surfaces sharing one context-bundle layer, 30+ FastAPI endpoints.
Reports, chat, insights — grounded in the same history
Three independent surfaces, one shared retrieval layer. The same context-bundle abstraction backs the report pipeline, the chat assistant, and the daily insights feed — different bundle profiles, one source of truth.
Four focused prompts, not one giant one
Summarise
Reads the user's structured context bundle — daily check-ins, behaviour segments, rolling aggregates, prior reports. Outputs a tight thematic summary, not a draft. The job is what to write *about*, not how.
- context bundle (full)
- report-type config
- thematic summary block
- candidate angles
- One prompt that tries to summarise *and* draft does both badly. Splitting the summarisation step lets the model focus on shape before voice.
Summary
A US-based behavioural-coaching SaaS was shipping charts and tables; users couldn't read their own data. LeadLyft is the AI services layer behind it — generating long-form personalised reports, a daily insights feed, and a chat assistant grounded in each user's history. Designed the architecture, built v1 end-to-end as the only engineer, then continued through the AnalyzeEQ document-intelligence module before being promoted to PM / Tech Lead. In production today with paying customer organisations.
Highlights
- Designed the 30+ API surface across reports, chat, and insights
- Multi-step Claude pipeline (summarise → analyse → draft → revise) per report type
- Reusable context-bundle abstraction shared across all three surfaces