When AI Agents Declare Victory Too Early
My AI orchestrator said "verified" on checks that could never have failed. Why AI agents declare victory too early, and the one question that catches it.
Arseniy Potapov
AI Agent Engineer
I build AI agents that work in production, backed by 16 years of Python.
Featured writing
All postsWhy Big Rewrites Fail
What 25 years of case studies reveal about the rewrite decision - from Netscape's fatal mistake to how Twitter, Facebook, and Shopify got it right.
Using Claude Code Without Technical Debt
Battle-tested workflows for using Claude Code in production. Two review modes and context-priming techniques that catch AI code bugs before they ship.
Selected work
Browserious
I build hosted browser sessions for agents. Browserious gives agents a real browser to navigate sites, interact with pages and inspect their work with screenshots.
MaestroPrivate beta
I build an agent orchestrator that coordinates coding agents across tasks. It gives them isolated work environments and carries their work through review and verification.
Testimonials
During a time when Host Compliance saw significant growth and profitability, Arseniy proved to be a valuable member & leader of the engineering team. He could always be counted on for his technical acumen, thoughtful system design, and mentorship. I can confidently say that Arseniy is one of the finest engineers I've ever hired and I would love the opportunity to work with him again.
Arseniy is a great developer capable of solving complex problems. In our years of working together he would consistently achieve high quality results and was someone team members would look to when they needed help. Any team is lucky to have him.
More testimonials6
Arseniy is simply one of the best developers I've had the pleasure of working with. We relied on his abilities greatly when it came to the most complicated tasks in our system.
Arseniy is an excellent Python developer and highly recommend him for any software development team. He is an excellent team player with loads of experience developing highly scalable web applications and big data processing systems in AWS.
It was real pleasure working with Arseniy- he was an integral part of our development team at Crossrider. He took part in building one of the most fast-growing mobile ad marketplace systems in the world, processing billions of transactions daily.
Ars worked on the slackbot, email automation and serverless related tasks for us. He is very skilled and we love working with him. We look forward to working with him in the future.
Arseniy is a great developer and very good at communicating!
Arseniy has been the BEST contractor on Upwork I've worked with. He's a great team member and a great asset to success of any project he touches.
Background
Résumé (PDF)Senior Software Engineer - AI & Data Engineering
TrueRoll builds software for local governments and property assessors.
I build an agents platform for business workflows and the harnesses coding agents work in: their context, tools and quality gates. I design CLIs that agents and engineers can both drive. I also build document AI for extracting and reviewing structured data, and code-review automation, and I connect these systems to production services with the observability they need.
My work also includes data pipelines, API integrations and cloud infrastructure.
Full experience8 earlier roles
Senior Full-Stack Developer
Market Making as a Service. LQFI provides unified digital asset trading infrastructure, enabling high velocity, efficient, and secure trading. Crypto liquidity provider on centralized and decentalized exchanges.
Being one of the core engineers in the algo-trading team
- I developed algo trading strategies for Binance / Kraken / Huobi / FTX
- Integrated cold wallets with the platform via Web3
- Extended internal framework, added multiple API integrations
Senior Backend Engineer
The World's #1 provider of short-term rental compliance monitoring end enforcement solution for local governments. Host Compliance makes it easy for municipalities to implement and enforce fair and effective short-term rental rules. SAAS for the US government to control short-term rentals, compliance, reports and tax collection. Integrations with third party services, web scraping, data processing, payments, auth, public APIs. Granicus acquired HostCompliance in 2020, I continued my role as a software engineer.
- I developed a platform API for short-term rental platforms to comply with the city of Los Angeles laws. This had a significant effect on cleaning fraudulent/unregistered listings from Airbnb and other platforms in LA.
- Created an API for user file uploads with a focus on security and. FedRAMP guidelines. This feature allowed encryption, very granular access control, and tight security policies to be implemented.
- Developed part of the platform allowing property management companies to renew their rental permits, report and pay taxes in bulk rather than individually.
- Built a basement for GraphQL back end for a new dashboard, which was highly modular and flexible and used a schema-first approach. This allowed the back-end team to extend the features quickly and kept the front-end team less dependent on back-end releases.
Sr. Tech Consultant
Platform & dashboard for industrial plants to control and optimize their energy consumption. Multi-level access, email & phone notifications, automated report generation, machine learning prediction & segmentation, direct plant-to-cloud data streaming integration. Focus on scalability, availability & cost-efficiency.
- Consulted client regarding algorithms, performance issues, tech stack, and solution architecture. Researched when required. Played a leading role in building a long-term strategy, solution architecture designing, planning sprints, grooming the task backlog, and preparing the tickets.
- Managed project and agile development team. Lots of code reviews done. Mentored and onboarded new developers. Tracked team's KPIs and goals.
- Handled all dev ops work and configured the cloud infrastructure from scratch.
Co-Founder & Sr. Tech Consultant
Managed a team of senior developers working on various projects, such as energy consumption monitoring technology, money remittance platform based on crypto, advertisement campaigns auto-optimization, mobile ads exchange, etc.
- Built an algorithm for the auto-optimization of internet advertising campaigns, which maximizes their performance by finding and blocking traffic sources and platforms that do not convert well enough.
- Architected and implemented ETL for a mobile advertising affiliate network. I used proprietary technologies for predicting the likelihood of conversion based on device data.
- Developed a bot detection feature that redirected bots from website content to other destinations—protecting content from crawling.
- Implemented email marketing automation and CRM integration, using SparkPost API and Zoho CRM.
- Consulted clients in the tech area, architected solutions, and managed an agile dev team. Participated in a lot of planning, communication, brainstorming, and architecting. Tracked team's KPIs and managed the goals.
- Played multiple roles - team lead, lead developer, scrum master, project manager, code reviewer, QA, mentor
Software Developer
Worked on automating all manual processes. Completely based on serverless approach, fully controlled via Slack bot.
Integrated with various APIs: Mailgun, Slack, Google Spreadsheets,
AWS & DevOps Consulting Services. I worked on internal projects for automating weekly reports generation: from gathering data from multiple sources and picking up notes from the relevant team members to data analysis & visualization and actual delivery via email & web.
Senior Developer
Worked as core engineer in personal website construction service, helped it evolve into a remote job platform and professional social network with 2M users online. Besides other tasks I solely worked on machine learning algorithm for smart job matching.
Integrated branded.me with outsource.com. Joined databases, doubled user base, added job matching and other services. Work time tracker app.
- Developed social networking features (follows, likes, and feeds) to be almost linearly scalable.
- Spotted and successfully coped with several bottlenecks in the back-end API code and infrastructure; made huge speed and robustness improvements.
- Wrote an algorithm to match users together based on contact lists users upload (a social networking feature).
- Performed email service integration, including dealing with bounces, unsubscribes, campaigns, and conversion funnel (marketing-specific experience).
- Integrated the billing, which includes work with user's accounts, transactions, adjustments, and recurrent payment plans (experience with user's money).
- Applied machine learning (NLP, clustering, neural networks, classification) to match users with open job positions based on their profile and job description. Also developed an algorithm to predict the required skills for a job based on its description.
Senior Python Developer
Core developer, worked on a high load distributed service serving millions mobile ads per second.
- I developed a high load distributed service hosted on Amazon AWS (EC2, S3), based on Django, using Python, Redis, MySQL, HP Vertica, Twisted, Celery, gevent and a little bit Lua on top of Nginx for lower level networking.
Fixed a system to scale linearly by profiling and eliminating critical bottlenecks, and dramatically reducing IO usage on each cluster node. Implemented a global lock feature on top of a Redis DB. Whenever it should be guaranteed that the shortcode section has exclusive access to some portion of data, it just acquires a lock, and other instances won't be able to reach it until release. Dramatically reduced latency by adding memorization on top of a local process memory and 2-level caching (local and remote Redis). Integrated with Google Ads, MobFox, and other RTB partners (real-time bidding). Created a special filtering layer on the level of Nginx. Lua scripts apply all the filtering efficiently by dropping not-useful connections very early, even before it gets to the Python back end.
Senior Python Developer
Business Services, Custom Software & IT Services. I was a developer team lead, developed web portal for cold call tracking & marketers training. Fully managed development process.
Developed a platform for the salespeople where they could schedule meetings, record their progress, and where they can self-educate to sell better and increase their KPI. Managed a team of developers, discussed features with the client, planned sprints, and then distributed tasks to developers, did showcases, and gathered feedback. Used Agile methodology. Developed a data-set preparation layer for a cloud machine learning platform: the data was accepted in different formats, and it has to be unpacked, parsed, shuffled, normalized, split onto parts, and fed into the machine learning workers' queue. Created a data set management system: a user can upload his data sets, review them, view some statistics, configure how to process them, download and remove them. It then passes it into the machine learning session OFC.
Expertise
- AI Agent Systems
- Harness engineering, orchestration, context, knowledge bases, quality gates and CLIs for agents
- LLMs in Production
- Document AI, RAG, evals and observability4 years
- AI & Machine Learning
- PyTorch, computer vision, fine-tuning and datasets4 years
- Python
- FastAPI, PyTorch, Pandas, NumPy16 years
- Data Engineering
- ETL, Kinesis, data pipelines14 years
- AWS
- SageMaker, ECS, Kinesis, Bedrock14 years
- Databases
- PostgreSQL, vector databases19 years
- Web Development
- FastAPI, Vue, Django, TypeScript16 years
- Docker
- Containers, ECS12 years
- Linux
- Ubuntu, Debian19 years
Tools and methods
- Agents and LLMs
- Agent Orchestration, Multi-Agent Systems, Harness Engineering, Context Engineering, Knowledge Bases for Agents, MCP, Claude Code, Codex, Evals, LLM Observability, OpenTelemetry, LLM, RAG, Prompt Engineering, OpenAI API, Anthropic, Gemini API
- Machine learning
- AI/ML, Document AI, Computer Vision, OCR, PyTorch, Transformers, Fine-Tuning, Dataset Building, CVAT, AWS SageMaker, AWS Bedrock, Vector Databases, Qdrant, ChromaDB, Ollama, llama.cpp, MLX, ONNX
- Backend and data
- Python, FastAPI, Django, Data Engineering, SQL, PostgreSQL, Data Modeling, Redis, Elasticsearch, API Integrations, Distributed Systems
- Infrastructure
- AWS, Docker, Linux, GitHub Actions, Datadog
- Frontend
- TypeScript, React, Vue
Education and certificates
- B.S. in Software Engineering, Energo University (ex. AUES)
- Machine Learning, Coursera
- Advanced Track of Introduction to Artificial Intelligence, Udacity
- Top Developer 2015, HackerRank
More projects7
TaktPrivate beta
I build task management and a knowledge base for agents, so they can find context, track dependencies and keep what they learn available for the next task.
play-tts
I build a local text-to-speech and speech-to-text server for generating audio and transcribing speech.
Simple OTP Demo
Simple OTP is a functional and secure solution demo. It allows users to verify their email or phone number by entering a unique One-Time Password (OTP) code. The demo app was built using a combination of cutting-edge technologies, including React, Material-UI, and AWS Lambda, DynamoDB, and the Serverless framework.
With Simple OTP, users can enter their email or phone number, receive an OTP code, and then enter the code to receive a signed JWT token that contains their email or phone number in the sub claim. This demo app demonstrates integration with third-party APIs such as Twilio and Mailgun for SMS and email services, respectively. It showcases use of Python, PyOTP, PyJWT, PynamoDB, and Requests libraries for server-side programming and API integration.
Push To Notion (web app integration)
Push to Notion" is a tool that allows users to quickly and easily add items to their Notion to-do list from various sources, such as Siri, Alfred, Slack, and Telegram. The motivation for this project came from the need for a convenient way to add items to a Notion to-do list without having to open the Notion app and locate the appropriate page. This can be especially useful for users who are on-the-go or occupied with other tasks, as it helps to maintain momentum and context.
@scan_driver_license_bot (telegram)
The ID Scanner Telegram bot is a reliable, efficient tool for extracting information from ID cards and driver licenses (US only). By sending a photograph of the document to the bot, users can receive the parsed data in JSON form as a response.
This bot utilizes the capabilities of AWS Lambda and Amazon's computer vision API to accurately process and extract data from the image, without storing any information or images. As a stateless, serverless service, the ID Scanner Telegram bot offers a secure and private experience for users.
Try the ID Scanner Telegram bot to streamline the process of extracting data from physical documents.
@PageScreenshotBot (telegram)
The Webpage Screenshot Telegram bot allows users to easily capture full-height screenshots of web pages. Simply send a command to the bot with the desired URL, and receive a screenshot as a response.
This bot leverages Puppeteer and a containerized Chrome instance to generate high-quality screenshots, all powered by AWS Lambda. With the Webpage Screenshot Telegram bot, capturing and sharing website content has never been easier.
Try out the Webpage Screenshot Telegram bot today and simplify your web content capture needs.
Match Cards (game for kids)
an interactive educational game using React for the frontend and Python for scripts and API integrations. The game helps kids learn English vocabulary through a fun and engaging experience, where players match spoken words with corresponding pictures. The media used in the game includes unique, custom-generated images, thanks to Midjourney AI, making the game more visually appealing and immersive. The game is a unique and effective educational tool for young learners, with its focus on language learning, interactive design, and cutting-edge technology.