Taif Ahmed Turjo

Hi, I'm Taif Ahmed Turjo. I'm doing my B.Sc. in Computer Science and Engineering at the University of Dhaka. Right now I'm a founding engineer at Rokkhe, where we're building an AI system that runs penetration tests on its own — it maps the target, finds a way in, proves each finding with a working exploit, and opens a pull request with the fix.

Before Rokkhe I worked on document checking for IPO prospectuses and fund factsheets at Makebell, and trained and reviewed code for language models at Scale AI, Turing, and SuperAnnotate. On the side I research retrieval: my project FedCRAG looks at how dense retrievers forget older documents as corpora grow and change, and how small per-client adapters can stop that without anyone sharing their raw documents.

Email / CV / LinkedIn / Github

Research Interest

Most of my questions sit where retrieval meets continual learning. In a RAG system the retriever is usually trained once and then frozen, but real corpora keep moving — new filings arrive, guidelines get revised, and different clients hold different collections. Keep fine-tuning a dense retriever on new batches and it slowly loses its grip on older material; in a federated setting, averaging adapters from clients with disagreeing corpora makes things worse. I want to measure that forgetting honestly — retention matrices, backward transfer, checks that rule out optimizer noise — and then prevent it with per-client LoRA adapters and aggregation that accounts for how severely each client forgets. I'm drawn to the same idea elsewhere too: agents that have to prove what they claim, whether that's an exploit for a security finding or a citation for a research report.

Education

University of Dhaka

B.Sc. in Computer Science and Engineering

2022 - Expected Graduation: February 2027
CGPA: 3.57/4.00

Work Experience

	Founding Engineer, Rokkhe June 2025 - Present, Remote Building the autonomous pentesting engine: multi-agent workflows over a 60-module skill library, a verification loop that runs each finding's proof-of-concept in a Docker sandbox before drafting a remediation PR, and a FastAPI control plane with live agent telemetry.
	Software Engineer, Makebell October 2025 - May 2026, Remote Built the document-intelligence and fact-checking pipeline for IPO prospectuses and fund factsheets, then reworked it into parallel multi-agent runs that cut processing latency by 70%.
	AI Trainer & L10 Reviewer, Scale AI December 2024 - January 2026, Remote (Contract) Fine-tuned and reviewed technical code across C, C++, Swift, Python, JavaScript, and Java, and designed adversarial tasks that expose where models break on multi-step reasoning and tool use.
	Rust LLM Engineer — PR Writer V2, Turing September 2025 - November 2025, Remote (Part Time) Reviewed and improved Claude's generated Rust as part of Anthropic's PR Writer V2 effort — correctness, error handling, library habits — and reported the recurring failure patterns.
	Python AI Engineer & QA, SuperAnnotate June 2024 - Present, Remote (Contract) Wrote training data for code models — snippets, bugs, and fixes in Python, C, C++, and HTML — and 500+ answers to programming questions used to train a chatbot.

Achievements

Champion, NSUCEC Cybernauts CTF
June 2026 — Inter-university capture-the-flag competition, NSU Campus

12th Place, University of Asia Pacific CyberSiege
May 2025 — Inter-university capture-the-flag competition

Projects

AURORA — Auditable AI Research System
Rust / code

An 8-crate system that takes a research question through plans, retrieval tasks, and evidence checks to a grounded report. Every claim carries a SHA-256-pinned source snapshot and a resolved citation, so a report can be rebuilt and its statements traced back to exactly where they came from.

Rust, OpenAI, Z.AI GLM, Tavily, structured output validation

Hopper — VM Provisioning Platform
Cloud Infrastructure / code / live

A platform for provisioning student VMs: a Go/gRPC orchestrator on Kubernetes with a FastAPI gateway and SvelteKit console, persistent per-user workspaces on replicated Longhorn storage, credit-based billing, and an admin operations suite — with a 590+ test suite in CI.

Go, FastAPI, SvelteKit, gRPC, Kubernetes, Longhorn

Cody — Collaborative Coding Platform
Real-Time Web / code

A real-time collaborative IDE: projects and file management, invitations with role-based access, and WebSocket presence across projects — with the groundwork laid for CRDT co-editing and Docker-sandboxed execution.

Next.js, FastAPI, CockroachDB, WebSocket, Docker

Awaaz — Women's Safety App
Mobile / code

A safety app for women: SOS alerts that go out in under three seconds, GPS tracking with safety-zone analysis, and a community feed where verified users post and help each other.

Flutter, Firebase, Supabase