About
Surafel Sentayehu
I grew up in Addis Ababa, Ethiopia, where I discovered programming through a desktop computer as a child. That curiosity led me to graduate from Addis Ababa University with a 3.96 GPA in Computer Science, train intensively at Africa to Silicon Valley (A2SV, backed by Google), and earn recognition from a Google Staff Engineer as a "fantastic engineer and true full-stack contributor."
Today I build AI-powered products at Manifest Automation, where I've helped grow a website to 45K visitors and 1.3M+ impressions in three months.
Outside of code, I play the Begena, a traditional Ethiopian 10-string lyre. It keeps me grounded. I build side projects with my twin brother and love video games.
Degree
Currently
Training
Outside work
Experience
Jan 2024 - Present
Full Stack Engineer
Manifest AutomationSunnyvale, CA (Remote)
- Built a full-stack SaaS console using React and Django for AI chatbot services, adding multilingual support, data filtering, and a custom chatbot builder that improved performance by 30%.
- Delivered 231+ tracked issues over 28 months across the React Native app, Django backend, Wagtail CMS website, and admin console, shipping 14 production mobile app releases.
- Led a full brand and domain migration (BibleChat → CrossTalk) across backend, mobile, and web, including production cutover, redirects, and Google Search Console coordination.
- Re-architected the real-time chat system using WebSockets, adding reconnection, typing indicators, connection status handling, and background-aware push notifications.
- Rebuilt the mobile onboarding flow end-to-end in React Native, boosting completion rates to 54.8% and install-to-trial conversion from 1% to 3.22%.
- Built the Customer Experience Portal (CEP), an agentic AI system that routes marketplace buyer messages into Slack and drafts seller responses using OpenAI and Model Context Protocol (MCP).
- Built content moderation systems including LLM-based blog spam filtering, a shadow-ban layer, and AI vision-based image moderation.
- Drove internationalization and SEO using structured data and Redis caching for a Wagtail CMS site that reached 45K visitors and 1.3M+ impressions in three months.
- As team lead, drove 9 production releases over 4 months by owning sprint planning, issue delegation, and QA oversight in Linear for a 3-engineer team, shipping features and resolving incidents.
Sep 2023 - Sep 2024
Full Stack Engineer
AppDiv System DevelopmentAddis Ababa, Ethiopia
- Built the frontend for a smart fleet management system (1,000+ government vehicles) using React.js + Material UI.
- Integrated WebSockets for real-time GPS tracking, handling 500K+ data points/day, 20% faster incident response.
- Designed fuel consumption algorithms on the backend (Jakarta EE), improving accuracy by 30%.
Dec 2022 - Oct 2023
Web Developer Intern
Africa to Silicon Valley (A2SV)Addis Ababa, Ethiopia
- Led 60% of the frontend for A2SV's website clone using Next.js, React, TailwindCSS, and TypeScript.
- Completed intensive training in Data Structures, Algorithms, and interview preparation, backed by Google.
Oct 2021 - Oct 2022
ML Engineer & Software Developer
4k LabsAddis Ababa, Ethiopia
- Built AI and robotics projects using Python and C++, including computer vision and machine learning systems.
Projects
Things I've shipped.
All projectsFaith-based mobile app serving 100K+ users across 190 countries with AI Bible chat, community prayer wall, Bible reader, clip sharing, and multilingual support across 30+ locales. Shipped 14 production releases.
Django monolith powering crosstalk.ai: Wagtail CMS with SEO Bible content, Prayer Wall, multilingual blog, and internal admin console for managing AI assistants, conversations, and analytics. 45K visitors and 1.3M+ impressions in 3 months.
An immersive real estate platform where each listing becomes a 3D digital twin generated from floor-plan geometry at runtime. Explore properties in first-person, step inside in WebXR VR, or place a scaled model in your real room via AR on mobile. An AI concierge powered by Gemini handles natural-language search, property Q&A, and instant mortgage estimates.
Digital garage management for EV service centers in Ethiopia: vehicle intake, repair tracking, Kanban status board, diagnosis & cost estimation, and a manager KPI dashboard. Built for EV Garage.
Professional portfolio website for an architectural firm showcasing interior design, exterior design, renovation, construction, and finishing work. Features bilingual Amharic/English support, email contact integration, and Sanity CMS for easy content management.
Arsenal fan card generator. Create a FIFA Ultimate Team-style card celebrating the 2025/26 Premier League title. AI-generated roast bios via Google Gemini, bilingual English/Amharic UI, and client-side MP4/GIF/PNG export.
Modern landing website for Ethiopia's premier Orthodox musical instrument training institute. Built to attract students, increase tutorial book sales, and facilitate donations while showcasing 10+ years of expertise.
SEO-optimized company website with integrated careers portal, automated job application system, and intelligent candidate filtering for a California-based respite and home care service provider.
Multi-tenant Telegram catalog bot for Ethiopian boutique owners. Turns product photos into an organized storefront where customers browse by category, filter by price, and order via a shareable Telegram link.
Mobile practice coach for the Ethiopian Begena (10-string lyre). AI-powered feedback via signal processing, using MFCC analysis and onset detection. Built with React Native and custom C/C++ audio modules.
Research
Per-Class Coverage for Rare Classes in Federated Conformal Prediction under Differential Privacy
Conformal prediction turns a classifier into sets of labels that contain the true label at a chosen rate, and federated learning trains a model without moving client data. Existing private federated conformal methods only hit that rate on average, so on long-tailed data the rarest classes, like a fraud pattern each bank has seen a few times, can be covered far less often. PC-FCP pools before it adds noise: each client counts its calibration scores into one histogram per class, secure aggregation releases their sum with Laplace noise added once, and the server reads a separate threshold for every class.
- Noise is added once inside secure aggregation, so it does not grow with the number of clients and rare classes are calibrated on their count across the whole federation.
- The full release is differentially private under a single budget; without noise every class keeps the label-conditional coverage guarantee, and a margin variant keeps a proved bound under noise.
- The only private method that holds per-class coverage at every budget from ε = 1 to 8: worst class covered 90.0% on CIFAR-10-LT and 85.3% to 85.9% on CIFAR-100-LT, while DP-FedCP drops to 0.0% at ε = 1.
Skills
Programming Languages
Frameworks & Libraries
Development and AI Tools
Achievements
Google Recognition
2026Recognized by a Staff Engineer at Google as a "fantastic engineer and true full-stack contributor."
Local 3rd Place in a global hackathon using NASA open data to tackle real-world challenges.
A2SV Alumni
2023Completed Africa to Silicon Valley's intensive DSA program. Backed by Google. 50+ alumni at Google, Palantir, Bloomberg.
LeetCode Annual Badge
2023Earned annual LeetCode badge for consistent problem-solving and competitive programming.
Scrum Fundamentals Certified
2023SFC certification. Proficient in Agile Scrum methodology and iterative software development.
Certificate of participation in CodeMaster, a coding competition for sharp, passionate engineers.
Certified for assisting students worldwide in Python programming training.
Completed the Elements of AI online course by the University of Helsinki.
Endorsements
I had the pleasure of working with Surafel for just under a year at Manifest Automation, and I would absolutely work with him again without hesitation.
Surafel is a fantastic engineer and a true full-stack contributor. He consistently demonstrated a strong ability to tackle a wide range of technical problems, from backend systems to frontend work, and did so with reliability and care. What really stood out to me was his curiosity and drive to learn — he was always exploring new technologies and patterns and proactively sharing those learnings with the rest of the team.
He's extremely hardworking, dependable, and brings a great attitude to collaborative problem-solving. Any team would be lucky to have Surafel, and I'm confident he'll continue to do excellent work wherever he goes.

Ian Talarico
Tech Lead @ Google
Manifest Automation · same team · Feb 2026
"I had the pleasure of working with Surafel at Manifest Automation on several projects, where he consistently demonstrated not only strong technical skills, but also great initiative and professionalism. I would gladly work with Surafel again."
Manifest Automation · different teams · Feb 2026
"I had the pleasure of working alongside Surafel at Manifest Automation, and I was consistently impressed by his ability to take full ownership of complex projects. He didn't just write code — he drove results. He has a rare talent for balancing robust backend architecture in Django with a seamless user experience. Surafel is the kind of engineer who bridges the gap between technical execution and business goals. I highly recommend him to any team looking for a proactive and skilled developer."
Manifest Automation · different teams · Feb 2026
Contact
Let's work together.
Open to remote roles, freelance projects, and interesting collaborations. Drop a message and I'll get back to you.









