← All work
LLMRAGPineconeNestJSReactForecasting
AI-assisted delivery forecasting
An internal planning tool that uses LLMs and semantic search over past work to estimate effort, predict delivery delays and surface productivity issues for the leadership team.
- Context
- Internal planning tool at a B2B marketplace
- My role
- Hands-on builder
- When
- 2025–2026
- commits I wrote across backend & frontend
- 600+
- re-estimation on every Jira change
- Real-time
- for the management
- 1 dashboard
The challenge
The management had to decide what to build next without a reliable answer to “when will it actually be done?”. Estimates were inconsistent, and delays showed up too late to react.
What I built
- Similarity-based estimates. When a ticket is created or changed, a vector search finds similar completed tickets. Their real effort patterns are fed to an LLM, which produces an estimate.
- Estimates at every level: epics, subtasks, research and design work, with AI estimates as the fallback in the delivery timeline.
- Always up to date. Jira webhooks trigger re-estimation
- Decision dashboards for the management: utilisation, builder performance, and AI analysis of each work package.
- AI-assisted intake. New requests are pre-filled by AI analysis, and semantic search finds related work instead of relying on exact keywords.
- Production-grade LLM plumbing: model routing with fallbacks and tracing of every LLM call.
Why it matters
Forecasting is where AI quietly pays off. It doesn’t replace judgement; it gives leaders a consistent baseline and an early warning, so the conversation shifts from “why is this late?” to “what should we cut?”.