Freight CRM

My contribution

• React SPA framework — converted the team's static HTML mockups into a React 19 + Vite app and set up the routing, sidebar, context, and the fetch wrapper that every later bucket landed inside.
• Buckets 6B–11 — Email Drafts pipeline (generator + UI + scheduler), Enquiry Board with Stage Tracker + Timeline, Sea-freight forms (FCL/LCL imp/exp), Air-freight forms + mode picker, Agent Dispatch with the 4-hour reminder loop and IMAP auto-reply threading, and the Rates table with L1 freeze + margin calculator.
• SPA integration for Buckets 2–5 — merged my teammates' AI scoring, field extraction, reply parser, and SMTP sender modules into the React app in a single merge sequence on 2026-05-25.
• Documentation — sole author of REMAINING_WORK.md and BUCKET_7_9_ASSUMPTIONS.md; co-author and primary maintainer of the developer guide.

What it does

End-to-end enquiry-to-quote pipeline. Gmail IMAP polls the client's ops inbox → an AI scorer ranks each new message → high-scoring messages auto-create enquiries, mid-scoring go to a review queue → field extraction lifts mode, ports, weight, and incoterm → the enquiry lands on a six-column Kanban → agents are dispatched over SMTP for rates → replies are parsed (PDF / Excel / text) into structured rate rows → quotes are built with a margin calculator that hard-locks below the threshold → L1 freeze snapshots the chosen rate and advances the stage with an immutable audit row.

Architecture

A Python IMAP sidecar (gmail_proxy.py) polls Gmail every 60 seconds and POSTs to a secret-guarded ingest webhook (INGEST_SECRET) on the Express backend, decoupling inbox polling from the request path. Express runs against a single SQLite file in WAL mode (31 tables, indexed foreign keys); the React SPA is served as static assets from the same Node process behind Nginx, with TLS via Certbot and PM2 keeping the process alive under systemd. A nightly cron snapshots the database to a GCS bucket.

AI scoring (Bucket 3, integrated)

Five weighted signals score each incoming message — sender domain (30), subject keywords like quote / RFQ / freight rate (25), body terms + quantity + incoterm (20), PDF/Excel/packing-list attachment (15), first-in-thread bonus (10). Thresholds: ≥75 auto-creates an enquiry, 50–74 routes to a review queue, <50 silently skips. Tunable through a sandbox UI on the Inbox screen — Shreshta built the scorer; I wired the sandbox into the React SPA.

Margin calculator + L1 freeze (Bucket 11)

The margin tab computes Buy(INR) = Buy(USD) × exchange_rate (default 83.50, mutable in settings). A slider sets the margin 0–30%; under 12% the route hard-locks the quote with a "needs approval" badge. L1 freeze is an idempotent endpoint — it snapshots the INR conversion using the current settings.exchange_rate, advances the enquiry stage, and writes one immutable row to enquiry_timeline. Re-firing the route doesn't double-snapshot.

Team & work split

Built in a team of four — a lead who brought the team together and three developers. My buckets are listed above; crediting the rest by name and bucket below, partly because it's accurate, partly because portfolio-grade transparency on freelance team work is the framing reviewers actually want.

• Mohith — team lead; connected the team and ran the engagement end-to-end. Repo bootstrap, branch integration / release management, production deploy runbook (GCP / Nginx / Certbot / PM2), Bucket 6C (Gmail App Password UX). I owned the Node-side production build-serve config (Vite build pipeline + static-middleware ordering) the runbook deploys against; GCP / Nginx / TLS were lead-led.
• Shreshta — Buckets 3, 4, 5 and 6A (AI scoring, field extraction, reply parser, SMTP sender) — the ingest layer I integrated into the SPA.
• Lochan Sai — Buckets 12, 13, 14, 15 (quote versions, analytics, agent rating + Best-Fit ranking, Rate Recommendations, LLM-backed reply analysis, lost-deal patterns) + ops layer (jobs / customs / documents tables).

The team shipped 15 buckets in total; my scope was Buckets 6B–11 plus the SPA framework and the integration of Buckets 2–5 into it.