Problem: sales, marketing and management ran on disconnected sheets and chat threads — no single source of truth.
Architecture: pipeline state machines with guarded transitions, role-based access down to field level, append-only audit log, revenue attribution engine joining touchpoints to deals.
Outcome: one canonical record per client, every follow-up traceable, management reporting in real time.

Problem: product knowledge lived in senior staff's heads; new salespeople took months to ramp.
Architecture: retrieval-augmented generation over chunked SOPs and transcripts, tool-calling into live CRM records, Bangla/English mixed-language prompts with guardrails against hallucinated pricing.
Outcome: instant, grounded answers for the sales team — in the language they actually speak.

Problem: sales die in chat threads — slow replies, no follow-up, quotes without decisions, staff drowning in inboxes.
Architecture: multi-intent understanding with scoped conversation memory; retrieval grounded on 1,200+ knowledge chunks with content-hash freshness; quote lifecycle with approval gates, immutable snapshots and inventory holds; automation engine for follow-ups with opt-out, cooldowns and quiet hours; structured handoff packets when a human must take over.
Evidence: text customer journey live-verified (guest send → AI answer in ~4s → persisted history); 82-test suite green; 73 migrations repeat-safe; human escalation preserved by design, never silently cleared.

Problem: field visits were unverifiable — paper logs, no proof of presence, no live visibility for managers.
Architecture: tracking runs only on explicit versioned consent + active check-in + registered device. Points queue in SQLite first, then upload as idempotent batches with exponential backoff; the server validates range, accuracy, freshness and impossible movement. Live positions fan out over Socket.IO; RBAC scopes every read and every location access writes an audit log.
Outcome: QA PASS — 105 web tests, 90 behavior/security gates green, signed production APK (com.harrington.field v1.2.2). Offline queue verified: 11 queued points drained to zero on reconnect with original timestamps intact.

Problem: field visits logged on paper; client records duplicated across reps.
Architecture: one-client-one-catalog enforcement via MongoDB multi-document transactions, GridFS-backed private photo storage, signed HTTP-only session cookies, full audit trail on every mutation.
Outcome: consistent client registry, photo-verified visits, management reporting without manual reconciliation.

Model: the client operates the platform under their own brand and sells it to Union Parishads on subscription + credit — union-wise wallets, invoices and transaction accounting built in.
Flows: citizen registration → application with file upload → status tracking → officer/UDC verification → approval → QR-coded certificate with unique number and approval log.
Scope: 47 services (certificates, citizenship, inheritance, trade license and more), reseller onboarding, package pricing and sales reporting.

Problem: a mock-data prototype needed to become a real system handling real money.
Architecture: JWT auth with role-gated admin routes, double-entry-style ledger tables, seeded member/settings data migrated to persistent MySQL, every financial mutation wrapped in transactions.
Outcome: a working ERP the cooperative can actually be audited on.

Problem: citizen services buried in an unmaintainable legacy site.
Architecture: Liferay DXP content types with approval workflows, REST-first service layer, Elasticsearch for Bengali/English search, WCAG 2.1 AA compliance throughout.
Outcome: 2M+ monthly visits on a portal citizens can actually use.