Production Work

What I've Built

5 production systems. Real scale. Real constraints. Each one started with a schema, not a mockup.

01
Authnex
Centralized Identity & Access Management Platform
TypeScript Node.js Redis Cloudflare OAuth 2.0
+
  • Centralized auth authority serving all internal SaaS products — replaced per-product login fragmentation with unified identity infrastructure.
  • JWT + rotating refresh tokens — each token invalidated on use, preventing replay attacks. Access tokens intentionally short-lived.
  • Multi-tenant RBAC via junction-table architecture — one user, multiple orgs, distinct roles per tenant. Zero account duplication.
  • Hybrid session layer — Redis for sub-ms blacklist lookups + MySQL for durable persistence. Full-user session kill on suspicion.
  • Local JWT validation by consuming services — stateless, horizontally scalable, no auth bottleneck per request.
  • Cloudflare deployment with IP anomaly detection, rate limiting, OAuth 2.0, social login.
// Architecture Flow
Client → Cloudflare Edge → Authnex API → Redis Blacklist
                                                + MySQL (durable)
Consuming Service → validates JWT locally (stateless)
Stack
TypeScript · Node.js
MySQL · Redis
Cloudflare Workers
JWT · OAuth 2.0
Highlights
Multi-tenant RBAC
Token rotation
Stateless validation
Blacklist caching
Status
Production · Active
02
AI Interview Engine
Autonomous Candidate Evaluation System
LLM Orchestration STT / TTS Node.js OpenAI · Gemini
+
  • 30s → <2s latency — root cause: tight sequential STT→LLM→TTS coupling. Decoupled pipeline, async execution, parallel API orchestration.
  • Plugin-style model abstraction — runtime switching between OpenAI, Gemini, Claude, Mistral, ElevenLabs, self-hosted. Zero code change per swap.
  • Prompt orchestration with role-specific templates + incremental transcript context. Structured JSON output enforced with schema validation.
  • Hybrid scoring engine — LLM eval + rule-based validation + weighted aggregation + confidence filtering. All async via concurrent workers.
  • Extensible test engine — MCQ, subjective, AI-evaluated formats. Role-based dynamic test generation.
  • Retry + fallback routing — provider failure triggers automatic reroute to next in priority chain.
// Pipeline Architecture
Audio Input → STT (async) ⟳ LLM (parallel) ⟳ TTS (async)
                              ↓
                       Scoring Workers → Evaluation Report
Models
OpenAI · Gemini
Claude · Mistral
ElevenLabs
Self-hosted
Key Win
30s pipeline
→ <2s end-to-end
Status
Production · Active
03
HRMS Platform
Multi-Tenant Enterprise Workforce Management
PHP MySQL Redis LAMP
+
  • End-to-end workforce platform — recruitment, onboarding, attendance, leave, training, certifications, PMS, incentives, MIS reporting.
  • Hierarchical org modeling — parent-child department schema, cycle-prevention constraints, structural integrity enforcement.
  • Concurrency-safe attendance — DB-level UNIQUE(employee_id, attendance_date) prevents duplicates. 20% fewer discrepancies.
  • Multi-reviewer PMS — weighted dynamic scoring, configurable aggregation, management override, evaluation locking.
  • Rule-based incentive engine — formula-driven payouts tied to PMS. Rules as DB records = zero-code updates.
  • Indexed views + query tuning → ~25% reduction in MIS analytics latency.
Stack
PHP (LAMP) · MySQL
Redis · jQuery
PHPMailer · Bootstrap
Modules
Recruitment · Attendance
Leave · Training
PMS · Incentives · MIS
04
Helpdesk & SLA System
Multi-Tenant Ticketing with Full Email Engine
PHP IMAP / SMTP MySQL CDN
+
  • Full IMAP integration from scratch — inbound ingestion, threading, reply handling, archiving, attachment management.
  • 3-layer deduplication — Message-ID + In-Reply-To (primary), subject normalization for Re:/Fwd (secondary), body validation (tertiary).
  • Contract-based SLA engine — timezone-aware UTC deadline computation, 24×7. Immutably attached at ticket creation.
  • Attachments on object storage + CDN — decoupled from DB, served from edge.
  • Debugged complex concurrent IMAP duplicate suppression in production under load.
Hardest Part
IMAP thread
reconstruction +
3-layer email
dedup in prod
05
DocXtractor
Hybrid AI Document Extraction Pipeline
Multi-model AI OCR Async Workers
+
  • Multi-model abstraction — Gemini, OpenAI, Claude, Mistral, self-hosted vision models. Same plugin pattern as AI Interview Engine.
  • Hybrid extraction — LLM semantic parsing for unstructured content + regex/rule-based deterministic validation for structured fields (PAN, GSTIN, invoice numbers).
  • Document-type templating — classified on ingestion, routed to per-template prompt + validation schema.
  • Structured JSON normalization at pipeline exit for reliable downstream integration.
  • Async concurrent workers — bulk uploads parallel-processed. Synchronous threads never blocked.
Approach
LLM parsing
+ regex validation
= hybrid accuracy