Ahmed Muharram
AI/ML Engineer based in Cairo, Egypt. I build production tool-calling LLM
agents, retrieval-augmented generation (RAG) systems, and high-throughput
backend infrastructure end to end, with a focus on systems and scientific
LLM agents.
About
I'm a recent graduate of the University of Pennsylvania (Bachelor's and
Master's in Computer and Information Science, May 2026) and an Onsi Sawiris
Scholar. I care about clarity, reliability, and the human experience of
software, and I work where systems and AI meet, bringing rigor to complex
problems while keeping people at the center. I build privacy-first AI
products and tool-calling LLM agents for scientific workflows, with a
background that spans research and industry.
Education
-
University of Pennsylvania, Master of Science in Engineering in Computer
and Information Science (Apr 2024 to May 2026).
-
University of Pennsylvania, Bachelor of Science in Engineering in Computer
and Information Science (Aug 2022 to May 2026), with minors in
Mathematics, Data Science, and Engineering Entrepreneurship.
-
Onsi Sawiris Scholar: a fully funded, merit-based scholarship awarded to
top Egyptian students for study at leading US universities.
Experience
-
Software Engineer (IT Applications Consultant), Orascom Construction (Jul 2026
to Present). Sole developer and technical lead on the rewrite of corporate IT's work-management system from ASP.NET WebForms to .NET 10, Angular, EF Core, and PostgreSQL, including migrating the legacy data from SQL Server. Designed, built, and owns the AWS platform that hosts the company's R&D applications (ECS on Graviton, RDS, CloudFront, WAF, CloudFormation), with reusable GitHub Actions deploy workflows that use OIDC.
-
AI/ML Engineer, Fidelity Investments (Dec 2025 to Jan
2026). Productionized and scaled an internal LLM/RAG system, expanding it
to arbitrary enterprise data sources via a standardized connector,
ingestion, and indexing template. Replaced a BM25S index with a custom
hybrid PostgreSQL full-text search + pgvector design, cutting refresh
latency from 10 minutes per 1M chunks to under 1 second; built
multivector and late-interaction retrieval with hybrid-score reranking
and a sharded Qdrant cluster. Now used in production by around 200 daily
R&D users.
-
Data Science and AI Intern, Fidelity Investments (Jun
2025 to Aug 2025). Architected a high-throughput ingestion pipeline for
terabytes of data at 99.99%+ success (schema validation, retries,
idempotent writes); built a grounded RAG application on Qdrant and
pgvector with hallucination checks and an evaluation harness;
containerized with Docker and Kubernetes.
-
Software Engineering Intern (AI Integration), Aydi (Jun
2024 to Aug 2024). Built a low-connectivity chatbot for agricultural
business owners (Haystack, MongoDB Atlas, Redis, GPT-4o) with 85%+ top-1
retrieval accuracy; reduced LLM spend by 96.25% via a cost and quality
model migration.
Master's Thesis (University of Pennsylvania, 2026)
How Underspecified Prompts Shape Tool-Calling LLM Agents in Scientific
Workflows, advised by Chris Callison-Burch and read by Eric Wong. Through two
open-source case studies on the Kani framework (LAMINA, a 50-tool
materials-feasibility agent, and PathFinder, a 44-tool pathogen-genomics
agent), it finds that prompt underspecification dominates model scale,
routing policy, and task complexity as a source of agent failure, with
instability localizing to the natural-language-to-tool-invocation mapping
rather than to tool execution. This work received the Penn Engineering
Outstanding Research Award.
Read the thesis (PDF).
Publications
-
Delip Rao, Ahmed Muharram, Chris Callison-Burch.
Same Claim, Different Interpretations: How LLM Agents Resolve Ambiguity in Scientific Claims. arXiv preprint arXiv:2610.20457, October 2026. Introduces Claim Cards, which make an LLM
agent's interpretation of an underspecified scientific claim explicit before it gathers
evidence, and shows that changing one interpreted field can flip the agent's verdict on about
a third of claims.
-
Ahmed Muharram. How Underspecified Prompts Shape Tool-Calling LLM Agents in Scientific
Workflows. Master's thesis, University of Pennsylvania, 2026.
Selected Projects
-
PathFinder: a 44-tool pathogen-genomics LLM agent on a
pydantic-ai and LangGraph pipeline (scoping, discovery, planning,
execution, verification) that compiles natural-language questions into
executable WDK strategy graphs across the 14-database VEuPathDB consortium
(PlasmoDB, ToxoDB, CryptoDB, VectorBase, and others), with pgvector
catalog grounding; now being integrated into the VEuPathDB platform.
https://github.com/ahmedomuharram/pathfinder
-
LAMINA: a 50-tool materials-feasibility LLM agent over
CALPHAD, Materials Project DFT, and CHGNet neural surrogates that turns
open-ended feasibility questions into traceable, executable tool
workflows. https://github.com/ahmedomuharram/lamina
-
ScenIQ: on-prem, security-first meeting intelligence
(SOC 2, GDPR, and HIPAA aligned) that turns meeting video into
transcripts, topics, and action items via FFmpeg, speaker diarization,
Whisper transcription, segmentation, and action-item extraction; FastAPI,
PostgreSQL + pgvector, Zoom/Teams ingestion, and ClickUp/Jira outputs.
https://sceniq.app
-
ByteNet: a distributed web search engine on a custom Java
stack (webserver, distributed key-value store, RDD-style framework) with a
~500k-URL crawler and distributed TF-IDF + iterative PageRank on AWS EC2.
-
InstaLite: a full-stack Instagram clone on AWS with
semantic profile search and real-time chat over WebSockets.
-
PennOS: a UNIX-like OS simulator with a preemptive
SIGALRM scheduler (priority queues) and a PennFAT (FAT16-style)
filesystem.
Skills
-
Agents and LLM frameworks: tool-calling and multi-agent systems, RAG,
prompt engineering, model routing, guardrails; pydantic-ai, LangGraph,
LangChain, Haystack, Kani, Vercel AI SDK; OpenAI, Anthropic, and Gemini
APIs.
-
Retrieval, embeddings, and LLM ops: pgvector (HNSW), Qdrant, BM25, hybrid,
multivector, and late-interaction retrieval, reranking, nomic-embed,
Whisper; OpenTelemetry, Langfuse, SigNoz; LLM evaluation (QWK, F1,
bootstrap CIs).
-
Backend and languages: Python, TypeScript, SQL, Java, C; FastAPI,
Pydantic, SQLAlchemy, asyncpg; REST, OpenAPI, WebSockets, Server-Sent
Events; distributed systems.
-
Frontend: React, Next.js, Vite, Tailwind, Zustand, TanStack Query,
ReactFlow.
-
Cloud and DevOps: PostgreSQL, Redis, MongoDB; Docker, Kubernetes, AWS (EC2,
RDS); GitHub Actions, CI/CD; Pytest, Vitest, Playwright.
Awards
-
Penn Engineering Outstanding Research Award, for the Master's thesis on
tool-calling LLM agents in scientific workflows.
- Onsi Sawiris Scholar.
Certifications
Writing
Essays and notes on LLM agents, retrieval, and building software, published at https://muharram.dev/#writing.
Links
- GitHub: https://github.com/ahmedomuharram
- LinkedIn: https://www.linkedin.com/in/ahmed-muharram
- Website: https://muharram.dev