Curriculum Vitae
Ahsen Tahir · Fredericton, NB, Canada · ahsentahir007@gmail.com
Graduating May 2027. Applying for thesis-based MSc programs in Canada for Fall 2027, available to start September 2027.
Education
B.S. in Computer Science, CGPA 3.55 / 4.00
- Final-year thesis: anti-distillation defenses that make capability theft computationally expensive, so one company's LLM cannot be cheaply copied into another's model by distillation.
- Relevant coursework: Artificial Intelligence, Information Retrieval, Probability and Statistics, Linear Algebra, Data Structures, Design and Analysis of Algorithms, Compiler Construction, Operating Systems, Software Design and Analysis.
- Dean's List: Fall 2023, Spring 2024, Fall 2024.
Aspire Leaders Program
Research experience
Mitacs Globalink Research Intern, AI + SE Research Lab, University of New Brunswick
- First author on Detecting Design-Level Security Anti-Patterns in LLM-Generated Microservices (with F. Palma, S. Khan), in progress.
- Built a catalog of nine design-level security anti-patterns from the microservice security literature, with explicit detection criteria and not-applicable conditions.
- Built an agentic detector that uses task-specific context modules, generates architecture diagrams, reconstructs service inventory, inter-service communication, deployment topology and auth/authz surfaces, and returns evidence-grounded verdicts.
- Validating the detector on 20 developer-built repositories with two independent human raters and a third resolving disagreements; next, using the validated pipeline to analyze microservice systems generated by Claude, OpenAI Codex, Google Antigravity CLI and Microsoft Copilot.
SASD: Source-Aware Self-Distillation
- Training-time defense against prompt injection, built on a multi-teacher KL divergence objective.
- Reproduced and benchmarked existing prompt-injection defenses on AgentDojo, establishing baselines across prior methods; the proposed approach already beats them, with ongoing experiments extending to privacy leakage and calibrated refusal. Work in progress.
AI Engineer
- Built agent-to-agent communication infrastructure used in production multi-agent systems, including structured protocols for task orchestration.
- Evaluated interoperability across multiple LLM agent frameworks: prompt design, dynamic role delegation, and structured output pipelines.
- This work led directly to the two publications below.
Publications
Anemoi: A Semi-Centralized Multi-agent System Based on Agent-to-Agent Communication MCP server from Coral Protocol
NeurIPS 2025 Workshop on Language Agents and World Models (LAW), 2025
arXiv:2508.17068
Beyond Rule-Based Workflows: An Information-Flow-Orchestrated Multi-Agents Paradigm via Agent-to-Agent Communication from CORAL
arXiv preprint arXiv:2601.09883, January 2026
arXiv:2601.09883
Other experience
Co-Founder
- Co-founded the company and led its pivot from a services agency to a product-led startup building OnClickAI, an embedded copilot layer for B2B SaaS products.
Deputy Head, Automation Team
- Built a WhatsApp chatbot handling student queries for PROCOM'25, with context retention across turns, FAQ resolution, PDF delivery, and fallback to a human representative.
- Added real-time intervention for organizers and query logging for analytics.
Content Writer (freelance)
- Wrote long-form articles explaining technical developments for a general audience, technical writing practice that carries directly into papers and research communication.
Selected projects
See the projects page for detail.
- Trend-X-BTC: LSTM with multi-head attention for 30-day Bitcoin forecasting over market sentiment, on-chain activity, and macro signals.
- Eeko AI: Llama 3.2 Vision for crop pest/disease detection and YOLOv5 for weed identification, with Urdu speech support for accessibility.
- VisionPlay: YOLO + ByteTrack for player and ball tracking in broadcast football footage, with optical-flow camera-motion correction.
Skills
- AI for software engineering: security and quality of LLM-generated code, automated software analysis, microservice architecture evaluation
- LLM post-training: supervised fine-tuning, instruction tuning, preference optimization (DPO, RLHF), knowledge distillation and multi-teacher objectives, KL-divergence training objectives, parameter-efficient fine-tuning (LoRA, QLoRA), quantization, synthetic training-data generation
- AI safety & robustness: prompt-injection defense, adversarial and adaptive attack evaluation, security–utility trade-offs, secure-code evaluation
- Agentic AI: multi-agent orchestration, agent-to-agent (A2A) protocols, Model Context Protocol, LangGraph, LangChain, CrewAI, tool use, long-horizon task decomposition
- Evaluation: benchmark design, held-out and adaptive evaluation protocols, ablation design, LLM evaluation
- ML / DL: PyTorch, TensorFlow, JAX, Hugging Face, attention architectures, RAG, vector databases (Pinecone, Qdrant, FAISS), computer vision (YOLO, OpenCV)
- Engineering: Python, C/C++, SQL, TypeScript, FastAPI, Next.js, Docker, Git
Awards
- Mitacs Globalink Research Internship, 2026: fully funded international research placement in Canada.
- Dean's List, FAST NUCES: Fall 2023, Spring 2024, Fall 2024.
- 6th nationally, SSC-II FBISE, 2021.
- Runner-up, AI Nexus Ignite Challenge, ACM-FAST Karachi, 2023.
- Runner-up, GitWit Bugathon, 2025.