[9iu7y7-+*-/ Fikresilase W. Abebayew — Full-Stack AI Engineer \]';[lkmj bhv="preconnect" href="https://fonts.gstatic.com" crossorigin />
Skip to Content

About

AI engineer who independently takes ambiguous products from architecture to production. Built an agentic knowledge system adopted by 6,790 bank employees, using hybrid retrieval, reranking, and information-verification loops. Built an agent-orchestrated reporting platform that reduced a 26-person, four week process to four hours. Launched a generative-learning platform in 12 days, reaching 135 users in its first five days. Authored an arXiv preprint on context-efficient agents and shipped its open-source reference implementation.

Experience

  1. Jul 2025 — Present

    In a bank, a wrong answer costs more than no answer. Plain retrieval kept saying it did not know, or returning half of one when the full answer lived across several documents. Nobody querying a company knowledge base actually wants a document handed back to them, they want the synthesis. So I built it as an agent rather than a search box: hybrid retrieval and reranking to find the right material, and a verification loop before anything reaches the user.

    • Sole engineer across ingestion, indexing, retrieval, LLM orchestration, backend APIs, and admin tooling, over approximately 3,000 institutional documents totaling 1.2M+ pages.
    • Adopted by 6,790 employees, nearly 70% of staff, with monthly actives growing to 6,314 across 23,500+ conversations and 115,000+ messages.
    • Writing the annual report meant weeks of research across spreadsheets, CSVs, and internal documents, so one long prompt was never going to hold. I built an agent layer with memory that knows the report structure, searches whatever documents or web sources a section needs, and assembles evidence piece by piece.
    • It checks each section for completeness, fills the gaps it finds, and produced the 2025/2026 annual report in four hours instead of a 26-person, four week process, with every claim traceable to source for human review.
    • System architect on a federated bank-wide lakehouse intended to become the primary source of governed data.
    • Built nine-category evaluation pipelines for the bank's voice and text agents, covering grounding and behavior testing before anything reaches a customer, and run vLLM serving where quantization and multi-token prediction made inference 3.5x faster.
    • Built integration and end-to-end test suites, GitHub Actions CI/CD, and RED-method observability with Prometheus, Grafana, and OpenTelemetry across retrieval, inference, and API services, deployed on RHEL inside the bank’s private data center.
    • Evaluate and prepare deployment readiness for an omnichannel customer-service agent still in development.
    • RAG
    • LangChain
    • FastAPI
    • vLLM
    • Weaviate / Milvus
    • Next.js
    • PostgreSQL
    • Docker
  2. Jan — Jun 2025

    • Led design and delivery of a comprehensive ERP system from scratch within a six-engineer team, delivered in six months and currently used as the ERP for the entire company with 2,400+ employees.
    • Built document-management, quality-audit, and operational modules with Next.js, Spring Boot microservices, and PostgreSQL.
    • Owned system architecture and application-wide development; Spring Boot services tested with JUnit 5, Mockito, MockMvc, and Testcontainers, Next.js with Vitest and Playwright.
    • Next.js
    • Spring Boot
    • Microservices
    • PostgreSQL
    • TypeScript
  3. May — Oct 2024

    • Joined as a research intern conducting independent research and prototyping in information theory, image and video compression, and sketching algorithms.
    • That work went on to inform my RBMQ image-compression preprint and final-year thesis.
    • Moved onto the engineering team to build the front end of a new content-management system in jQuery.
    • Worked on page performance, publishing workflows, and media-compression functionality.
    • Python
    • NumPy
    • OpenCV
    • Pillow
    • JavaScript
    • jQuery
  4. Jul 2023 — Jan 2025

    An e-signature platform for car dealerships, and the role where I learned that the hardest engineering problems are usually operational. I was the developer, the person on the customer call, and the one paged when something broke, which is a fast way to learn what actually matters in a codebase.

    • Migrated the frontend from a single unstructured HTML file to a componentized React application while the product was live with 1,200 active users, without service disruption. Integration and end-to-end test coverage is what made that safe.
    • Designed APIs and service boundaries across the Go and Node.js backends and maintained them for over a year at 99.99% uptime.
    • Fixed a production failure where client retries after a timeout caused duplicate processing: the client could not tell a failed request from a lost response. Made the affected endpoints idempotent with per-operation idempotency keys so retries return the original result instead of doing the work twice.
    • Ran roughly 50 client calls a month across onboarding, integrations, escalations, and requirements gathering, and fed that directly into what the team built next.
    • Maintained GitHub Actions CI/CD and RED-method monitoring with Prometheus, Grafana, and OpenTelemetry; deployed on Google Cloud.
    • React
    • Node.js
    • Go
    • Vitest / Playwright
    • GitHub Actions
    • Prometheus / Grafana
    • GCP
  5. Feb — Jul 2023

    This was never really an image problem. In early 2023 the hard part was getting reliable structure out of a model, before agent frameworks existed and before models were good at following a schema. Rather than forcing strict structured output, I had the model emit sign words: when the parser sees one, the tool performs the matching action. That was our own answer to agent orchestration and output structuring, at a point when neither had a name yet.

    • Co-founded and independently built a TypeScript/Electron desktop agent that renames, tags, and organizes image libraries from natural-language instructions.
    • As the sole technical member, took end-to-end ownership of architecture, agent behavior, OpenAI integration, reliability logic, and packaging.
    • Designed the multi-step agent and validation logic that turned organization intent into safe, consistent file operations.
    • Generated approximately $10,000 in total revenue from approximately 150 designer customers.
    • TypeScript
    • Electron
    • OpenAI API
    • Node.js
  6. Feb 2022 — Feb 2023

    A software consultancy where clients contracted us to build their products. I was on the build team and also the maintainer and customer-facing engineer for the accounts I worked on, which meant owning the consequences of my own code.

    • Built the entire application for CareYaya, a US care-and-education platform, on Vue, Nuxt, Hasura GraphQL, and Go, including GraphQL schema and permission design.
    • Maintainer and escalation point for client projects including Mastercard and the World Bank, handling roughly four technical tickets and three client meetings a week.
    • Practiced test-driven development with Vitest and Playwright on the Vue/Nuxt side and Go’s standard testing package on services; shipped through GitHub Actions to Google Cloud.
    • Vue / Nuxt
    • Hasura GraphQL
    • Go
    • Vitest / Playwright
    • GCP

Projects

  • Bestie, AI visual learning platform

    Bestie, AI Visual Learning Platform

    Bestie started as a research question: what is the most compute and storage efficient way to teach something visually? Video models are the obvious answer and the expensive one. So instead of generating frames, I designed a communication protocol between the model and a custom browser renderer, where the model emits a structured scene description and the browser draws it. That runs roughly 100 times cheaper and about 5 times faster than producing the same lesson with a video model.

    • Founder and sole builder. Researched the architecture for three to four weeks, then built and launched it in 12 days.
    • Designed a multi-stage narrator, director, and visual-animator pipeline producing versioned scene representations, with structured-output parsing, sanitization, repair, and fallbacks for reliability.
    • Built the rendering engine with React, SVG, GSAP, and KaTeX.
    • Shipped web, backend, worker, admin, and Electron desktop surfaces, including local MP4 export using parallel offscreen Chromium renderers and FFmpeg.
    • Instrumented the pipeline end to end. During the measured launch period every lesson completed, recovering automatically from a 6% scene-level retry rate, at roughly 54 seconds and $0.10 per lesson.
    • First five days: 135 users, 218 generated lessons, and 691 scenes.
    • Python
    • FastAPI
    • Next.js
    • TypeScript
    • PostgreSQL
    • Vertex AI
    • GSAP
    • Electron
    • FFmpeg
    • Docker
  • Sage, open-source LLM coding agent

    Sage, Open-Source LLM Coding Agent

    • Python terminal agent that analyzes projects and assists with understanding, refactoring, debugging, and modification, with OpenRouter model selection. Published on PyPI as sage-code.
    • Reference implementation of the Gatekeeper Protocol from my arXiv preprint, built around a low-fidelity structural map, progressive context requests, synchronized state, and declarative JSON actions to cut context cost and state drift.
    • The preprint reports 73% average task completion against 58% for the next-best RAG baseline, evaluated across three programming tasks and seven language models. It is an arXiv preprint, not peer reviewed.
    • Python
    • LLM Agents
    • Context Engineering
    • OpenRouter API
    • PyPI
  • Viral Clip, desktop AI clipping application

    Viral Clip, Desktop AI Clipping App

    OpusClip and the tools like it are expensive for what they do, so I built an open-source alternative and put it out. I never marketed it. It exists because I wanted the pipeline to exist, and the interesting part was the framing, not the clipping.

    • Windows desktop application that turns local or YouTube long-form video into short, shareable vertical clips.
    • Built the end-to-end pipeline in Python and PyQt6: source acquisition, subtitle retrieval or audio extraction, transcript chunking, AI viral-moment analysis, structured clip proposals, and overlap deduplication.
    • Implemented dynamic 9:16 reframing with MediaPipe multi-face and pose detection, optional YOLO object detection, target locking, dead zones, velocity prediction, and transition smoothing across scene changes to reduce jitter.
    • Added visual-analysis fallbacks for long videos and missing captions, caption generation, FFmpeg and MoviePy export, concurrent processing, and PyInstaller packaging.
    • Python
    • PyQt6
    • OpenAI API
    • OpenCV
    • MediaPipe
    • FFmpeg
    • yt-dlp

Research

  • The Gatekeeper Knows Enough, arXiv preprint

    The Gatekeeper Knows Enough

    Sole author · arXiv:2510.14881 [cs.AI] · October 2025

    • Introduced the Gatekeeper Protocol, a domain-agnostic framework for reliable agent and system interaction built on a low-fidelity latent context map, a synchronized state ledger, progressive contextualization, and declarative JSON actions.
    • Positioned the protocol as an alternative to increasingly complex agent-memory systems. Agents reason over an inexpensive structural map and request high-fidelity context only when they actually need it.
    • Evaluated five context-management strategies across three programming tasks and seven language models. Sage reached 73% average task completion (±8%) with 0.8 grounding errors and 6,200 total tokens, against 58% for the next-best RAG baseline.
    • Released Sage as the open-source reference implementation. This is an arXiv preprint, not a peer-reviewed publication.
    • LLM Agents
    • Context Engineering
    • Evaluation
    • Python
    • arXiv
  • Reduced Bit Median Quantization, arXiv preprint

    Reduced Bit Median Quantization

    Sole author · arXiv:2409.13789 [eess.IV, cs.IT] · September 2024

    • Proposed RBMQ, a middle-process technique that introduces exploitable pixel-value redundancy into an image before conventional file-format compression is applied.
    • Explored quantizing and reducing pixel representation from eight bits to five for storage and transfer efficiency, targeting deep-archive and large-image-transfer scenarios.
    • Implemented the application side of RBMQ as my final-year thesis project, which received an A+.
    • Seven-page arXiv preprint, not a peer-reviewed publication.
    • Image Compression
    • Information Theory
    • Python
    • NumPy
    • Pillow

Awards