Resume
Summary
Technical founder and hands-on CTO with five years of building production LLM agents used by 10K+ people, leading teams of 20+. I've trained and fine-tuned open-source models, built private on-premise inference servers used by a law firm for confidential work, and defended shipped agents against attackers and users trying to manipulate them. That experience led to my independent research on whether social persuasion can sway agents that control money, now moving from behaviour (2,850 negotiations across five models) to model internals. I'm drawn to systems we don't yet fully understand, and to explaining them simply.
Research
Persuasion attacks on budget-holding AI agents / Independent
November 2026 · github.com/sjuery/AgenticPersuasion
- Tested whether social pressure alone, with no prompt injection, makes an agent holding a budget overpay: 2,850 scored negotiations across 5 models (Claude Opus, Sonnet and Haiku 5.5, DeepSeek V4.1 Flash, MiniMax M3), 50 runs per condition, permutation tests, $24 total.
- Finding: agents spot and resist recognisable tactics (commitment and pretexting flagged in 46–48 of 50 runs) but not price anchoring, which raised the best binding offer by $162–$411 on every model tested (p < 0.001).
- My prediction was the reverse (commitment strongest). Next, comparing base and post-trained versions of the same open-weight model, and reading and steering its emotion representations, to find out why numeric pressure goes undetected.
Experience
Sparkit (formerly Spaarkd) / CTO
May 2024 – Current · Dubai · Machine Learning Engineer, Dec 2023 – May 2024
- Joined the venture-backed startup at a pivotal stage to rebuild the platform from scratch, with founding-team equity. Led an engineering team of up to 10 (20+ contributors over 2.5 years) building an AI platform that turns prompts into manufacturable products: 9.6K registered users, 20K+ AI-generated designs and 230+ registered factories and suppliers.
- Built Gabi, a production multi-agent system (8.4K conversations and 7.5K design renders since mid-2026) that replaced our earlier pre-agent system, plus an MCP server with an OAuth 2.1 connector so users can design and price products from inside Claude, ChatGPT and other platforms.
- Hardened Gabi against adversarial use: web content passes through quarantined research agents before it reaches the orchestrator, and actions with real-world effects, such as placing production orders or using a customer's logo, wait for user approval. It held up against real prompt-extraction and price-negotiation attempts.
- Instrumented every LLM call with model, token and cost telemetry (about 820M tokens across Gemini, DeepSeek and Claude) to choose models on evidence: for some agents, a frontier model at 8.5× the cost gave no quality gain, so we kept the cheaper models and reworked the tools. Earlier, trained and fine-tuned open-source LLMs and image models until frontier models overtook what custom training could reach at our budget.
- Built an agentic MRP platform that assigns each order to a factory, has agents track production, and sends customers, factories, suppliers and the team daily digests; plus internal agents for SEO content, competitor-ad monitoring and AI video production.
- Built a testing harness that checks the whole platform before each deploy (280 database test files and 44 end-to-end mobile flows across web, iOS and Android), so bugs from reports can be fixed without hands-on intervention and non-technical teammates can ship changes.
MNNT / Co-founder & CTO
September 2021 – Current (part-time since May 2024) · Switzerland / USA / Dubai
- Led a team of 20+ (developers, artists, 3D modellers) building a tokenised game that reached 13K monthly active players; raised about $2M through a token sale and a private investor.
- Turned the game's characters into personal AI agents with persistent memory and hybrid retrieval (pgvector + full-text search) across web, iOS/Android and real-time voice: 1.2K users have held multi-turn chats, and two-thirds came back after their first week.
- Built a browser agent in which an LLM drives a real browser (AWS Bedrock AgentCore + Browser Use) to book, fill in forms and sign in on any website, with spoken approval in voice mode, saved logins encrypted with AWS KMS, and 2FA handoff.
- Cut daily LLM spend by about 98% with no drop in quality ratings from QA testers and 1,000+ users, by trimming the system prompt from 13.6K to 4.9K tokens, restructuring prompt caching and routing chats to different models by complexity.
- Piloting health agents with hospitals in the Philippines on wearable data (heart rate, HRV, SpO2, temperature, sleep, steps, estimated blood pressure), for about 600 users so far. Built with school and hospital launches in mind: zero data retention on the chat model route, crisis detection that escalates to local hotlines, and fencing so injected content is treated as untrusted.
Hint3rland / Co-founder & CTO
January 2022 – May 2024 · Switzerland
- AI and Web3 consulting and R&D for clients including LVMH (Hennessy), Pangaia, Condé, HEAD and Dorier/MCI.
- Built private on-premise LLM servers with RAG over a law firm's documents, so confidential files never left the firm; fine-tuned open-source models for narrow use cases.
- Hennessy Chinese New Year capsule (2023–24): AI-generated zodiac art per customer plus collectible tokens, bridged between a Chinese permissioned chain and Ethereum. Worked with LVMH legal in Europe and China, filed the ICP licence and shipped the live site; the client shelved the launch during the crypto downturn.
- NFC-tagged caps for a racing driver's event, doubling as passes for holders of his earlier NFT collection.
- Sparkit started as a Hint3rland client before I joined it full-time.
Independent developer / Web3
January 2021 – January 2022 · Remote
- Built free crypto tools (earnings, copy trading, payments, tax calculators) that peaked at about 10K daily active users.
- Turned that audience into a paid development business: about 100 clients, 6–7 long-term, and a 1,000+ member Discord; hired developers to keep up with demand.
AdGuerillas Technologies / Software Engineer
May 2020 – January 2021 · Zurich, Switzerland
- DevOps and software for interactive advertising hardware (smart TVs, screen backpacks, holographic fans), including interactive games and usage analytics for clients.
Game developer / Fullas Games, MBBD
2016 – 2020 · London, United Kingdom / Switzerland
- Core gameplay and an AI opponent for a Hearthstone-style game at Fullas Games; lead game developer at MBBD; a Riot Games API web app built in a week for the Legends of Runeterra dev challenge.
Teaching
Coding for Lawyers 2.0 / Co-designer and co-instructor (Solidity)
2023 – Current · with lawyers Michel José Reymond and Fabien Gillioz, for the Digital Law Center and the
Ordre des Avocats de Genève
- Co-wrote and taught 6 sessions explaining smart contracts and their legal uses: small groups of about 10 practising lawyers, and university groups of 50+, including at the University of Geneva and at Pestalozzi Law in Zurich.
- A follow-up series on AI agents (how they work, their risks and uses in law) is being prepared.
- Started teaching at 15: taught older students to program video games to fund a trip to India.
Education
42 Silicon Valley / Software Engineering
September 2017 – May 2020 · California, USA
- Projects included a Dota 2 game-playing AI, A* pathfinding, an expert system and a ray tracer.
EPFL / Physics
September 2023 – May 2024 · Lausanne, Switzerland (left to join Sparkit)
Skills
AI/ML: multi-agent systems, tool use, MCP servers with OAuth 2.1, hybrid retrieval (pgvector + full-text),
fine-tuning open-source LLMs and image models (PyTorch, NumPy), on-premise inference (Hugging Face, vLLM, Ollama),
real-time voice and browser agents; Claude, Gemini, DeepSeek and MiniMax APIs, OpenRouter, Claude Code
Evaluation & research methods: controlled experiments (2,850 scored runs, 50 per condition), permutation tests,
per-call token and cost telemetry, choosing models from the data, prompt-cache and context-size optimisation
Infrastructure & data: GCP, AWS (ECS, Bedrock AgentCore, KMS), Docker, GitHub Actions, Vercel, Cloudflare,
Supabase, Postgres/pgvector, pre-deploy test harnesses, Ethereum and permissioned-chain bridging
Agent security: quarantining untrusted web content, approval gates for real-world actions, KMS-encrypted
credentials with 2FA handoff, defending against prompt extraction and price-negotiation attempts in production
Programming: Python, TypeScript, Node.js, Next.js, React Native, FastAPI, SQL, Solidity, C/C++, C#/Unity
Product & leadership: leading teams of 20+, client delivery with corporate legal teams (LVMH, ICP licence),
teaching technical topics to lawyers and university groups
Languages & interests
Languages: French (native) · English (fluent) · German (basic) · Russian (learning)
Interests: Travel (lived in ten countries) · Tennis · Cooking · 3D printing