Mikiyas Assefa

I'm a Computer Science student at Carnegie Mellon University in Qatar, graduating in 2028.

I build software with language models and do research on them. At Pronto I built the agent behind an AI shopping assistant that is live on their marketplace. At Nestlé I led a Qatar-wide shopper insights study and presented the recommendations to the Country Manager and leadership board.

At CMU-Q, advised by Eduardo Feo-Flushing, I co-authored a paper at ICASSP 2026 on using LLMs to correct speech recognition errors for people with aphasia, and TRAP-QA, a benchmark of common-sense questions with subtle traps (under review at KDD 2027).

Email CV GitHub LinkedIn

Experience

May – Aug 2025

Pronto

Applied AI Engineer Intern, Doha

Shipped a conversational AI shopping assistant into the production marketplace at pronto.qa (60+ designers, 2,000+ products).
Built the Python/LangChain agent backend with a FastMCP tool server: structured-output routing, RAG, complete-look retrieval, and session memory.
Enriched catalog metadata with LLMs (41 tags per product) and built a synthetic-user tester scoring 10 goals × 3 styles.

Aug 2026 – now

Carnegie Mellon University in Qatar

Computing Services Assistant

Tested CMU's campus LLM gateway (LiteLLM, vLLM) across 3 models, finding tool-calling failures and missing parser configs.

Aug 2026

Nestlé

Digital Strategy & Analytics Intern, Doha

Led a Qatar-wide shopper insights study; built a Python pipeline and interactive dashboard on buying habits, channel preferences, and retail conversion barriers.
Presented recommendations on visibility, conversion, and in-store execution to the Country Manager and leadership board.

Research

* equal contribution

Under review
KDD 2027

TRAP-QA: Benchmarking LLM Robustness Against Subtle Tricks, Traps, and Pitfalls in Cleverly “Silly” Casual Questions

Hangyi Wen*, Anas Semsayan*, Mikiyas Assefa*, Eduardo Feo-Flushing

KDD 2027 Datasets and Benchmarks Track, under review · code

294-item benchmark of common-sense robustness against deceptive traps, curated from ~6,500 candidates.
Built the evaluation pipeline (~12,000 LLM calls per run across 10 models) with a 3-judge panel and arbitration.

ICASSP 2026

LLM-Based Post-ASR Error Correction for Disordered Speech

Hangyi Wen*, Mikiyas Assefa*, Anas Semsayan*, Eduardo Feo-Flushing

ICASSP 2026, Barcelona · paper · code · poster

LLM error correction for aphasic speech, evaluated across 10 ASR systems (Whisper, AWS, Azure, GCP).
LLM fusion reached 14% WER (46% relative reduction); phoneme-matched few-shot retrieval gave a 53.8% relative reduction.

Projects

Oct 2025

IRRISmart

2nd place, Reboot the Earth Hackathon (United Nations) · code

Smart-irrigation platform built in 48 hours: a Random Forest forecasts soil moisture, temperature, and salinity 72 hours ahead, shown on a Mapbox farm map.

Apr 2026

Tides of Time: Al Jumail

demo · code

Interactive proposal site for revitalizing Al Jumail, a historic pearling village in Qatar, as a living heritage site.

Jan 2026

Mednova

demo

Offline-first emergency triage app that walks responders through adaptive, protocol-based questions to a triage outcome and next steps.

Teaching

Aug – Dec 2025

15-110 Principles of Computing

Course Assistant, Carnegie Mellon University in Qatar

Held office hours, led Python review sessions and recitations, and graded assignments and exams.

Awards

2024 – 2026

Dean's List · All-Star Student Merit Scholarship · SURA Apprenticeship · QSIURP Grant