Imam Chowdhury
Sr. Systems Engineer · Automation ArchitectI am Imam Chowdhury. I build and improve the infrastructure, security, automation and AI workflows that keep modern organizations running.
I'm Imam Chowdhury. I design reliable infrastructure and intelligent workflows for work that cannot stop.
The record, without the desktop
Sr. Systems Engineer · Automation Architect
I build and improve the systems that keep modern organizations running reliably, securely and efficiently. My work spans infrastructure, networking, security, automation and AI-driven workflow design — from disaster recovery, virtual environments, Microsoft platforms and hybrid mail systems to firewall, wireless and endpoint security.
§ 02The hall
The desktop above sits on a server hall. Each of the systems on this record is a rack in it, standing in the zone where its mission is played, and joined to the Core by two cables — one to each spine switch — so no single cable is a single point of failure. That property is computed from the drawing, not written about it.
Fabric AFabric BCross-connect
§ 03Flagship
“An AI editorial suite for reporters” is a line any CV can carry. This is what is behind it: 13 tools in daily use across The Daily Star's newsroom since Jun 2024, each built because a specific job was being done by hand. Editorial judgement stays with the editor throughout.
The route a story takes, each tool handing to the next.
Audio and video to timestamped text, chunked so a long recording does not defeat it. English and Bangla speech.
Before A press briefing or a leaked recording used to be transcribed by listening to it and typing.
A transcript becomes a publish-ready draft in house style, source-locked against invention, with the 5W1H checked before it is handed back.
Before Turning disorganised field notes into a structured story is slowest exactly when the news is breaking.
Grammar and house style in a tight or a minimal pass, without touching facts or quotes. Five headline options, paraphrasing three ways, theme regeneration — and the sub keeps every decision.
Before Basic copy-fixing was a bottleneck in front of everything waiting to be published.
Bangla and English in both directions, paragraph by paragraph, phrased as journalism rather than as machine output.
Before A bilingual newsroom was waiting on translation to move a major feature between its own two languages.
Meta title, description, keywords, URL slug and a readability score, derived from the finished article rather than guessed at before it.
Before Editors rarely have time to write metadata properly, and it is the part search engines read first.
Everything a finished story is turned into.
Reads a story, pulls the figures out of it, and returns a structured visual script the design desk can lay out.
Before Text reporters were the wrong people to be isolating data for graphics, and the graphics desk was waiting on them.
Captions and IPTC keywords written into the file itself, with location and entity tagging, ready to download.
Before Tagging was manual and went through Adobe Bridge; the metadata now travels with the photograph instead of beside it.
Platform-specific posts, pull-quote cards and hashtags, sized down from the story itself.
Before The digital desk was rewriting the same story by hand for each platform while the news was still warm.
Production-ready scripts generated from published articles in the channel's own style, with visual and audio cues marked.
Before Rewriting an article as a video script is a format change, and it was being done from scratch every time.
Subtitles for any video, in any language.
Before Subtitling by hand is tedious enough that it quietly stops happening.
The desks and the mornings the stories come out of.
Multi-page images and PDFs to organised text you can paste, including ePaper pages.
Before The archive's OCR route was manual and page-by-page.
Summarises the day across several publications' ePapers, follows jump pages to where a story continues, and sets the coverage against The Daily Star's own — what matched, what was missed.
Before Comparing the morning's papers by reading all of them is a job that has to finish before the day starts.
300+ CVs in a single pass, with an editable shortlist, CSV and PDF export, and candidate cards carrying strengths, weaknesses and comparisons.
Before Reading hundreds of CVs by hand is slow, and slow reading is where inconsistency gets in.
§ 04Systems
Fig. 4.1Newsroom Intelligence
Thirteen tools in daily newsroom use since mid-2024 — transcription, drafting, subediting, translation, SEO, photo metadata, OCR and an ePaper comparison desk — in one route from raw material to publication-ready copy. Editorial judgement stays with the editor throughout.
Internal system · live since mid-2024In the hall: Newsroom · Mission 02
Fig. 4.2Applied AI
A retrieval-augmented workspace over an organisation's own documents — ask a question, get an answer grounded in the source rather than in the model's memory.
In-house systemIn the hall: Cortex
Fig. 4.3Media Operations
The archive itself, rather than the tagging tool that feeds it: bulk ingestion for high-volume newsroom photography, with a datetime fixer that repairs the wrong-capture-date problem at the root and metadata embedded into the files themselves rather than held in a database beside them.
Operational systemIn the hall: Archive · Mission 03
Fig. 4.4Workflow Automation
The layer that chains the suite together for one story: ad graphics, accurate map documents, SEO groundwork and social cards produced in a single pass rather than as separate errands after publication.
Operational systemIn the hall: Forge · Mission 04
Fig. 4.5Workflow Automation
Repeatable document production that turns complex source material into consistent outputs.
In-house toolingIn the hall: Forge · Mission 04
Fig. 4.6Infrastructure
Active Directory, DNS, hybrid mail, virtualization and disaster recovery — the Microsoft and Linux foundations a newsroom that cannot stop is actually standing on.
Professional practiceIn the hall: Core · Mission 01
Fig. 4.7Cybersecurity
Next-gen firewall, switching and routing, wireless, and EDR aligned into one operational defence layer — including hardening the Windows and Linux privilege-escalation paths an attacker would actually take.
Professional practiceIn the hall: Perimeter · Mission 05
Fig. 4.8All of the above
Every system on this sheet is a rack in the hall drawn on sheet 02, dual-homed to two spine switches, its traffic routed over real paths. In the 3D Facility each one lights up as its mission is restored.
Enter the Facility§ 05Automation
The same way every time: find the step someone is doing by hand, put the simplest tool that does it reliably in its place, and keep the decisions with the person accountable for them. A rule where a rule will do; a model only where reading is actually needed; and when something new arrives that does the job better, learn it and use it.
Before An interview recording comes in every few minutes, and someone types each one out before a word can be written.
Machines do the typing and the first draft; the editor keeps the decision.
Before Thousands of photographs a day, half with the wrong capture date and none with tags. Finding one means scrolling.
Fix the root cause once, then let the cheapest model that is good enough do the rest.
Before Scanned forms are retyped into documents, checked by eye, and corrected again when someone spots a mistake.
A rule where a rule will do; a model only where reading is actually needed.
Before Stories now go out in English and Bangla. The line needs a translation step — and a new model has just been released.
Keep learning: a new tool, understood and put to work, made the line shorter and cheaper.
In the Facility these four lines are a puzzle: the machines on a bench, a clerk at a desk with the pile growing, and a conveyor to build. Put a step in the wrong order and the work jams where it stops fitting; hand a decision to a script and the line stops.Build them yourself
§ 06Timeline
Education, the three roles, every capability from the month it became daily work, what is running now, and the honours along the way — against the same years, so the overlaps and the growth can be seen rather than read.
§ 07Principles
Most paid tools solve a problem that an open-source stack solves as well, for the cost of understanding it. I would rather spend that understanding once than rent it forever — and I would rather say so than pretend the licence was the only option.
A solution that works and cannot be afforded has not worked. The interesting engineering is almost always in getting solid, recoverable behaviour out of what an organisation already owns.
The models and the tooling change every few months. So I favour systems with clear boundaries and readable seams — the kind an AI can help extend without a rewrite — over clever ones that only their author can maintain.
Most of what I run today I did not know how to run five years ago. Staying useful means treating unfamiliar problems as the job rather than as an interruption.