App Consultant
- Year
- August 12, 2026
- Role
- AI Engineer
- Client
- PT. Mitra Pinasthika Mulia
- Access
- Internal only
- LangGraph
- vLLM
- LiteLLM
- Next.js
- TypeScript
Internal system — no public link or screenshots.
What follows describes the problem, the approach and the engineering decisions at a level that discloses no proprietary detail.
App Consultant lets an engineer or analyst describe an application in plain conversation and walk away with architecture.md and prd.md already written to the company's internal standard.
The problem it solves is consistency. Every new internal application needed the same two documents, and every team wrote them differently, which made architecture review an exercise in reformatting rather than reviewing.
What it does
- Conversational intake: a guided chat draws out scope, constraints, integrations and non-functional requirements without the user filling in a form.
- Two generated artefacts: a product requirements document and an architecture document, both following the internal template.
- Compliance by construction: the templates encode the company's regulations, so a generated document starts compliant rather than being corrected into compliance.
- Local inference: runs against self-hosted models, so nothing about an unreleased internal system leaves the network.
Approach
The conversation is a LangGraph state machine rather than a single long prompt. Each stage collects one part of the specification and validates it before moving on, which keeps a long intake coherent and lets a user resume where they left off. Document assembly is a separate step from elicitation, so the template can change without touching the conversational logic.