Job Experience
I build backend systems, AI infrastructure, and enterprise applications. My work spans government-scale Enterprise systems, distributed data pipelines, and production ML workflows.
DYNAMIC SOLUTION INNOVATORS
Software Engineer | April 2023 – Present | Dhaka, Bangladesh
Core member of the engineering team building large-scale enterprise systems for government and public-sector clients, with a focus on backend architecture, security, and operational reliability.
Key Impact: Built and open-sourced gf, a GlassFish development CLI that hot-swaps bytecode into running JVMs via JDI/JDWP and bypasses JasperReports classloader caching, reducing local edit-deploy cycles from ~2 minutes to ~5 seconds.
E-Appeal: Digital Appeals System for NBR & USAID
A digital platform for managing tax-related appeals.
- Defined the security model, package structure, and CDI/EJB conventions for DSI’s first large-scale Jakarta EE system; later JEE projects at the firm adopted these conventions.
- Centralized RBAC enforcement by replacing authorization checks scattered across JSF views and services with a unified permission-evaluation layer.
SHMS: Smart Hotel Management System (built on Project Scratch platform)
A nationwide licensing and registration system for hotels and restaurants, developed on DSI’s internal core-services platform.
- Designed a reusable approval-workflow engine backed by configurable state transitions, replacing hardcoded per-module approval logic. Engineered a dynamic field-level correction system enabling reviewers to reopen only selected form fields with inline feedback while preserving audit history.
- Introduced centralized observability for GlassFish applications using Prometheus, Loki, and Grafana, exposing JVM memory, thread pools, and request latency for production debugging and capacity monitoring.
- Isolated auditing and notification dispatch from business logic using AspectJ-based cross-cutting interceptors, establishing a reusable pattern across multiple Jakarta EE systems.
BCIC-ERP: End-to-End Enterprise Resource Planning
A comprehensive ERP system for process automation across 12 integrated modules.
- Re-architected file storage and delivery behind a pluggable
StorageClientabstraction supporting MinIO and filesystem backends, replacing base64-inlined file rendering with URL-based streaming to eliminate large HTML payloads and reduce server-side memory pressure. Added scoped file authorization and nginxX-Accel-Redirectoffloading for efficient file serving. - Designed a JMS-based notification pipeline with DLQ-backed retries, isolating SMS/email failures from the synchronous request path.
- Led development of the Procurement, Inventory, Asset, and Budget modules of BCIC-ERP, including process modeling, schema design, backend implementation, and production rollout.
ALTRI.AI
Full-Stack AI Engineer, RevestAI (revestai.com) | Part-time, Remote (US)
Building AI-assisted real estate analysis systems that combine LLM-based enrichment, traditional ML valuation, and large-scale property ingestion pipelines.
Key Impact: Achieved ~10× throughput on LLM data enrichment by rebuilding the pipeline around bounded-concurrency asyncio.gather with atomic per-batch commits.
Core Contributions
- Designed the three-stage property scoring pipeline that separates LLM feature extraction, valuation models, and deterministic financial heuristics, so each layer can be retrained and tested on its own.
- Built the property ingestion pipeline on HomeHarvest and PostgreSQL with content-hash deduplication and incremental delta-syncing. Repeated scrapes no longer trigger duplicate enrichment runs, and historical listing state is preserved.
- Implemented distributed job coordination on PostgreSQL using
SELECT FOR UPDATErow-level locks with automatic stale-lock recovery, so multiple scheduler replicas safely process long-running enrichment and training jobs. - Built the ML training and promotion pipeline with Champion/Challenger validation. A challenger must beat the champion on a gold-standard holdout slice, and a human signs off before production-wide rescoring, since aggregate metrics can hide slice regressions and valuation errors cost money.