✓Copied to clipboard

Omik

I design declarative cloud environments, automate server provisioning, and keep Linux services online without manual toil. Whether writing modular Terraform stacks, tuning Prometheus alert rules, fine-tuning open AI models on domain datasets, or deploying autonomous triage agents into production daemons — everything I build is deterministic, observable, and version-controlled.

Omik — DevOps Engineer · Platform Automation & AI Systems
Bogura, Bangladesh (UTC+6)
INFRASTRUCTURE AS CODEDEPLOYED // PROD

Automated Multi-Node Cloud Platform with Terraform & Prometheus

Declarative Terraform HCL modules spinning up hardened Linux hosts, paired with automated Prometheus metric collectors and Alertmanager routing.

15s Alert Dispatch100% Declarative HCL0 Manual SSH Drift
TerraformPrometheusGrafanaAlertmanagerNode ExporterUbuntu LinuxHCL
Source ↗
AI MODEL TRAININGDEPLOYED // PROD

Domain-Specific AI Model Training on Custom Engineering Datasets

Fine-tuning open weights on internal documentation and system logs for accurate, zero-hallucination technical assistance.

< 0.1% Hallucination Rate4-bit AWQ Quantization65ms Inference Latency
PythonPyTorchLoRA / QLoRAHugging FaceFastAPIDockerJSONL Datasets
Source ↗
TECHNICAL COMPETENCIESVERIFIED IN PROD

Capabilities Matrix

Engineered without arbitrary skill percentage bars or vendor buzzwords. Tested operational tools, Linux kernel interfaces, and production automation frameworks.

Server Automation & IaC (Terraform & SaltStack)

Terraform Cloud & Bare-Metal Provisioning

HCL architecture, reusable modules (VPC, compute, security groups), remote state locking with S3/DynamoDB, and clean rollback paths.

SaltStack Configuration Management

Declarative SLS state formulas (wp-saltstack), pillar secrets, and automated configuration convergence across bare-metal and cloud nodes.

Linux Kernel & System Administration

TCP socket window tuning (somaxconn), cgroups v2 resource slicing, systemd service units, and automated firewall rules via nftables.

POSIX Shell Automation

Defensive Bash scripting (set -euo pipefail) for cold-start bootstrapping, zero-dependency server provisioning, and automated backup rotation.

Observability & SRE (Prometheus & Grafana)

Prometheus & Exporter Ecosystem

Node Exporter, cAdvisor, Blackbox probes, custom application metric instrumentation, and tuned scrape intervals to minimize server overhead.

PromQL & Alert Rule Engineering

Writing accurate PromQL queries for CPU throttling, memory headroom, disk I/O wait times, error budgets, and high-percentile latency distributions.

Alertmanager Routing & Deduplication

Grouping rules, route trees, notification dispatch to Slack/Discord/webhooks, and silencing flapping alerts during maintenance windows.

Grafana Production Dashboards

Golden signal dashboards (traffic, errors, latency, saturation) designed for rapid troubleshooting rather than confusing visual noise.

AI Model Training & Specific Datasets

Dataset Preparation & Curation

Cleaning, tokenizing, and structuring custom instruction-tuning datasets (JSONL, Q&A pairs) from technical documentation, logs, and git histories.

Model Fine-Tuning (LoRA / QLoRA)

Parameter-efficient fine-tuning on open foundation models (Llama, Mistral, Qwen) to achieve domain-specific technical accuracy and eliminate hallucinations.

Retrieval-Augmented Generation (RAG)

Vector embeddings, hybrid BM25 + semantic similarity search, and source-grounded prompt pipelines using vector stores.

Inference Optimization & Serving

Model quantization (GGUF, AWQ), running local inference engines (vLLM, Ollama), and building low-latency asynchronous FastAPI microservices.

Production Server AI Agent Integration

Supervised Server Agent Daemons

Running agents as supervised systemd services listening to Alertmanager webhooks and system signals with least-privilege security controls.

Automated Incident Triage Pipelines

Agent gathers non-destructive diagnostic context (ss, dmesg, journalctl, memory saturation), correlates the root cause, and drafts the runbook.

Docker & Zero-Downtime Rollouts

Multi-stage minimal image packaging, healthcheck dependency probes, blue-green container swaps, and automated rollback upon failed deploy.

Server-Side Tracking Pipelines

Server-Side GTM containers, Meta Conversions API (CAPI), and GA4 Measurement Protocol with Redis deduplication to preserve conversion telemetry.

CURRENT TELEMETRYLIVE

Now Ledger

Automating multi-node VPS provisioning with Terraform, training custom LoRA models on engineering logs, and deploying self-healing Linux triage agents.

eBPF tracepoints in the Linux 6.x kernel, low-overhead Prometheus metric scraping at scale, and dataset distillation techniques.

Available for Remote DevOps Roles, Terraform Stacks & AI Agent Deployments — Bogura, Bangladesh (UTC+6) — collaborating with distributed engineering teams worldwide.

SERVER AUTOMATION & AI AGENTSDEPLOYED // PROD

Production Server AI Agent with Automated Incident Triage

A background Python daemon connecting Alertmanager webhooks to an autonomous AI agent for instant failure diagnosis on live servers.

< 12s RCA GenerationRead-Only Sandboxed ExecZero Auto-Mutations
Python 3AsyncIOsystemdPrometheus WebhooksBash AutomationClaude / OpenAI APIsLinux Diagnostics
Source ↗
INFRASTRUCTURE & ORCHESTRATIONDEPLOYED // PROD

High-Availability E-Commerce & CMS Cloud Platform (wp-saltstack)

Fault-tolerant, auto-healing LAMP/LEMP platform orchestrated via wp-saltstack behind active-passive HAProxy and Nginx layer-7 ingress.

99.98% Measured Uptime< 180s Recovery TimeVRRP Sub-Second Failover
SaltStackHAProxyNginxDockerKeepalivedMariaDB GaleraUbuntu Linux
Source ↗
DATA & CONVERSION INFRASTRUCTUREDEPLOYED // PROD

Server-Side Tracking & Ad-Ops Telemetry Pipeline

Custom server-side GTM proxy and event broker preserving conversion telemetry through Safari ITP and ad-blocker signal loss.

99.4% Event Match Quality+32% Recovered Attribution0 Client CPU Script Latency
Server-Side GTMGA4 Measurement ProtocolMeta CAPIRedisNginx IngressLooker StudioDocker
Source ↗
CONNECT

Great infrastructure starts with automated precision.

Available for Remote DevOps Roles, Terraform automation stacks, Prometheus monitoring pipelines, and production AI agent integrations.

INFRASTRUCTURE AS CODE

Automated Multi-Node Cloud Platform with Terraform & Prometheus

15s Alert Dispatch100% Declarative HCL0 Manual SSH Drift

Modular Terraform HCL declaring VPC, security groups, and cloud-init bootstrap scripts. Deploys Prometheus with tuned scrape targets, pushing structured alerts via Alertmanager to on-call webhooks with zero manual server intervention.

Setting up new servers was previously done via manual SSH sessions, forgotten firewall configurations, and inconsistent packages. When services degraded under load, the team only found out after customer complaints.

Wrote clean, modular Terraform configurations to provision VPS instances, private networking, and firewall rules in minutes. Automated the deployment of Prometheus and Node Exporter via cloud-init, streaming real-time system metrics to Grafana. Configured Alertmanager to notify on-call engineers within 15 seconds if disk I/O saturates or CPU throttles.

Opted for cloud-init over heavy agent orchestration to ensure zero-dependency bootstrapping on fresh Ubuntu nodes.

TerraformPrometheusGrafanaAlertmanagerNode ExporterUbuntu LinuxHCL
Inspect Terraform & Prometheus Configs
AI MODEL TRAINING

Domain-Specific AI Model Training on Custom Engineering Datasets

< 0.1% Hallucination Rate4-bit AWQ Quantization65ms Inference Latency

LoRA rank-16 adapters fine-tuned on curated JSONL instruction pairs from production logs. Quantized to 4-bit AWQ and served via async FastAPI container running behind reverse proxy with strict schema validation.

Generic foundation models routinely hallucinated proprietary API endpoints, outdated configuration parameters, and internal command syntax when queried by team members.

Curated and cleaned thousands of structured instruction-tuning pairs in JSONL format from internal technical documentation and commit histories. Fine-tuned an open-source model using parameter-efficient LoRA techniques, quantized the weights for fast inference, and wrapped the model in an asynchronous FastAPI service. Reduced technical hallucination to near zero on internal workflows.

Adopted parameter-efficient LoRA over full fine-tuning to cut VRAM requirements by 75% while keeping base weights modular.

PythonPyTorchLoRA / QLoRAHugging FaceFastAPIDockerJSONL Datasets
View Training Pipeline
SERVER AUTOMATION & AI AGENTS

Production Server AI Agent with Automated Incident Triage

< 12s RCA GenerationRead-Only Sandboxed ExecZero Auto-Mutations

Event-driven systemd daemon with locked-down seccomp sandbox. Correlates Prometheus Alertmanager JSON webhooks with non-destructive system telemetry (`journalctl`, `ss`, `dmesg`), generating deterministic incident summaries.

Waking up at 3 AM to SSH into servers, manually running dmesg, ss, and journalctl, and figuring out why a worker died costs valuable time during production outages.

Engineered an event-driven daemon running as a systemd service. When Prometheus fires an alert webhook, the agent safely executes non-destructive diagnostic commands over SSH, filters out sensitive data, analyzes the root cause, and posts an incident summary to Slack with the verified rollback or fix command ready for human sign-off.

Enforced strictly read-only diagnostics with human sign-off for remediations to prevent catastrophic automated cascading failures.

Python 3AsyncIOsystemdPrometheus WebhooksBash AutomationClaude / OpenAI APIsLinux Diagnostics
Inspect Agent Architecture
INFRASTRUCTURE & ORCHESTRATION

High-Availability E-Commerce & CMS Cloud Platform (wp-saltstack)

99.98% Measured Uptime< 180s Recovery TimeVRRP Sub-Second Failover

Active-passive HAProxy with Keepalived VRRP routing to horizontally autoscaled Nginx + PHP-FPM clusters. Database state synchronized across MariaDB Galera multi-primary nodes, fully managed via SaltStack state formulas.

Monolithic e-commerce instances on single droplets suffered unmonitored crash loops, cascade failovers during sales surges, and configuration drift across servers.

Engineered reproducible infrastructure state using SaltStack (wp-saltstack formula) with automated state enforcement across Debian/Ubuntu nodes. Ingress terminates at dual HAProxy load balancers running keepalived for virtual IP failover, distributing traffic to horizontally scaled Nginx + PHP-FPM workers with automated Let's Encrypt TLS renewal and sub-180s recovery time.

Chosen SaltStack masterless/minion architecture for instant convergence speed over complex container orchestration overhead.

SaltStackHAProxyNginxDockerKeepalivedMariaDB GaleraUbuntu Linux
Inspect wp-saltstack State Engine
DATA & CONVERSION INFRASTRUCTURE

Server-Side Tracking & Ad-Ops Telemetry Pipeline

99.4% Event Match Quality+32% Recovered Attribution0 Client CPU Script Latency

First-party Nginx reverse proxy routing client JSON payloads to isolated sGTM Docker containers. Redis handles event deduplication and timestamp hashing before broadcasting signed payloads to Meta CAPI and GA4.

Client-side tracking scripts were losing up to 35% of ad attribution and revenue telemetry due to aggressive ad-blockers, iOS privacy sandboxing, and Safari cookie limits, distorting ROAS calculations.

Engineered a hardened Server-Side Google Tag Manager (sGTM) cluster hosted on a custom first-party subdomain behind Nginx with Redis event deduplication. Client events send raw HTTP data payloads directly to the server container, which enriches, normalizes, and dispatches server-to-server calls to Meta Conversions API (CAPI) and GA4 Measurement Protocol with 99.4% event match quality and zero client CPU overhead.

Maintained an in-memory Redis deduplication buffer rather than persistent disk storage to maximize throughput during high-traffic ad surges.

Server-Side GTMGA4 Measurement ProtocolMeta CAPIRedisNginx IngressLooker StudioDocker
View Tracking Architecture