Ammar Tahir
Building the future of intelligent systems — from agentic code frameworks to graph-powered AI.
About
Innovative AI engineer and researcher crafting next-gen intelligent systems — from multi-agent code reasoning to graph-driven automation. I fuse deep learning, systems design, and creativity to turn complex problems into elegant, scalable solutions.
Work Experience
Google DeepMindML ResearchContract
ML Researcher Intern
- - Built WebGym to synthesize React web-app datasets from GitHub commits, enabling reinforcement learning signals and cutting task generation from one month to 22 minutes across 100 tasks.
- - Authored research using LLM token log probabilities and semantic entropy to improve test validity and coverage; paper submitted to EMNLP.
- - Delivered an end-to-end multi-agent framework with Gemma models for reasoning, code generation, debugging, and automated test creation.
KaleidoscopeFounding Software EngineerTech Lead
Founding Software Engineer
- - Developed KAI with a retrieval-augmented generation pipeline over a knowledge graph using fine-tuned AWS Bedrock and Claude to power dynamic knowledge-lake queries.
- - Designed a three-tier caching layer with React Query, Redis, and Neptune to deliver sub-second responses for 100k+ graph nodes, reducing AWS spend by 16% and increasing engagement by 20%.
- - Partnered with product and infrastructure teams to ship a fault-tolerant monitoring system using Prometheus and TimescaleDB that ingested 10,000+ metrics per second.
Education
University of Washington, Seattle
National University of Computer & Emerging Sciences
Skills
Projects
SSSP Gym — BMSPP vs. Dijkstra
Built an interactive single-source shortest path visualizer with step-by-step playback and state tables, implementing both Dijkstra and BMSPP algorithms for side-by-side comparison.
Asciinema Web Player
Delivered a React-based player for asciinema .cast recordings with in-browser uploads, playback controls, and checkpoint support for agentic debugging workflows.
Ransomware Detection in VMs using ML – NetApp
Led a research team to build an entropy- and Random Forest-based ransomware detector for VMDK images, reaching 97% precision, 99% recall, and curating a 500-sample dataset with QEMU and SleuthKit tooling.
Community Detection in Social Networks using GNNs
Developed a community-detection approach for large-scale social networks, fine-tuned AGNN models, and achieved 70.1% accuracy with a 0.492 NMI score.
KnowledgeGraphGPT
Created an open-source tool that turns plain text into knowledge graphs with GPT-4, boosting accuracy via link prediction and rule inference techniques.
Apache TinkerPop
Fixed a production bug in Apache Gremlin by tightening null checks within the traversal bytecode serializer.
Source-Free Domain Adaptation
Researched transfer learning techniques to boost sentiment classification accuracy for pre-trained BERT-family models in low-resource settings.
PageBook
Built a cross-platform collaborative document editor with realtime presence, shared cursors, and conflict resolution.
Essentials – Microsoft Imagine Cup Regional Finalist
Led a full-stack team to build a food-waste marketplace with React and Node.js, and fine-tuned ResNet50 on 30k images to reach 96.87% classification accuracy.
Thulla – Card Game
Implemented a multiplayer, fault-tolerant card game server with GenServer supervision trees and live websockets.
FileSystem
Prototyped a reader–writer lock table to study concurrency control strategies inside database file systems.
Snake Xenzia
Recreated the classic Snake experience in x86 assembly with keyboard input handling and smooth animation timing.