Building UrjaOps: an operating system for PM-KUSUM solar projects

A PM-KUSUM solar project depends on many parties at once — MNRE, the State Nodal Agency, the DISCOM, lenders and an EPC contractor. Projects rarely fail on technology; they stall on execution: a missed filing window, a bank guarantee that lapses unnoticed, payments to a contractor that run ahead of the work, or a request sitting on someone else's desk.

Developers were managing all of this in spreadsheets, inboxes and memory.

• Lifecycle workflows for all four components (A, B, C-IPS, C-FLS), with phases, completion gates and known pitfalls.
• A RACI directory mapping every step to the responsible agency, with overdue follow-ups surfaced automatically.
• A deadline register for guarantee expiries, commissioning and filing dates, with red/amber alerts and email reminders.
• Seven portfolio registers: PPA, land, bank guarantees, risk, legal matters, EPC contracts and consortium.
• An EPC tracker that flags when money paid runs ahead of certified work.
• 28 AI drafting actions — letters, checklists, SOPs and reviews — grounded in the project's own data and exported to DOCX.
• A state-level monitoring view for nodal agencies.

Next.js 14 and TypeScript on Vercel; Supabase Postgres with row-level tenant isolation; Razorpay for billing; Resend for email; Vercel cron for deadline reminders and weekly summaries; LLM drafting with prompt guardrails that mark every output as a draft for human review.

• Provenance on every figure — badges show whether data came from a document, a user or a template.
• Fail-closed sharing — read-only share links use 128-bit random tokens and deny on any error.
• Clean tenant data — demo data can never be written into a customer's workspace.
• Human in the loop — the AI drafts; people review and send.

UrjaOps is live in production with free project tracking and a public sample-project demo. It shows how we work: learn a regulated domain deeply, encode the workflow, and put AI where it saves real hours — with guardrails.