# Hector Oviedo > Senior Applied AI Engineer and AI Solutions Architect, 23+ years of software engineering. Agentic systems and multi-agent orchestration, MCP servers and clients, RAG, computer vision, multimodal generative AI, and local inference optimization. 1st place, Anna AI-Native App Hackathon 2026. Remote worldwide, based in Rosario, Argentina. Delivery is agentic engineering: coding agents produce the bulk of the code, while architecture, specs, integration tests, security review, and deployment stay human-in-the-loop. Every claim below carries a link that shows the evidence: a live production site, a public repository, a package registry entry, a press release, or the organizer's own announcement. Links checked 2026-08-11. Canonical site: https://hec-ovi.dev (non-www, single page, static export). ## Profiles - [GitHub, hec-ovi](https://github.com/hec-ovi): 50+ public repositories. Everything claimed here is readable commit by commit. - [GitHub profile README](https://github.com/hec-ovi/hec-ovi): long-form index of the same work with the same sources. - [X, @hec_ovi](https://x.com/hec_ovi): display name hec.ovi, bio "Senior Applied AI Engineer", links back to hec-ovi.dev. - [LinkedIn, hec-ovi](https://linkedin.com/in/hec-ovi): primary contact channel. LinkedIn serves HTTP 999 to crawlers, so this one is not machine-readable. ## Recognition, with sources - [Anna Hackathon 2026 winner announcement](https://x.com/Anna_Partners/status/2076694358817222743): posted 2026-07-13 by @Anna_Partners (verified account, "Anna AI"). Text: "1st Place: Gamentic by @hec_ovi". Nine finalists reached Demo Day, three placed. X serves this page only to signed-in clients. - [Anna, the platform](https://anna.partners): AI partner platform that ran the AI-Native App Hackathon. - [Anna developer hub](https://anna.partners/developers): documents Executas (JSON-RPC plugins an LLM can call), the format the gamentic port ships in. - [Anna AI-Native App Hackathon event listing](https://www.eventbrite.com/e/anna-ai-native-app-hackathon-tickets-1990936738379): online, June 15 to 22, 2026, organized by "Anna | Your Proactive AI Partner". - [First place certificate](https://hec-ovi.dev/anna-hackathon-2026-first-place.jpeg): image hosted on the site. - [ADGS awards page](https://www.adgs.com/awards): ADGS states PANDEXIT "was selected by the jury at the World Innovation Summit for Health 2020 as one of the most innovative health application worldwide". He engineered the PANDEXIT dashboard end to end. - [Press release, PRWeb, 2021-01-27](https://www.prweb.com/releases/deep-technology-startup-adgs-releases-pandexit-a-covid-variant-simulator-awarded-by-world-innovation-summit-for-health-894829283.html): same release, photo caption reads "Eng. Nahuel Gonzalez, Eng. Hector Oviedo and Christophe Billiottet". - [Press release, EIN Presswire, 2021-01-25](https://www.einpresswire.com/article/534798970/deep-technology-startup-adgs-releases-pandexit-a-covid-variant-simulator-awarded-by-world-innovation-summit-for-health): the release naming him as an ADGS engineer on PANDEXIT. - [CCI France Qatar, 2021-01-23](https://www.ccifq.com/actualites/n/news/deep-technology-startup-adgs-releases-pandexit-a-covid-variant-simulator-awarded-by-world-innovation-summit-for-health.html): the French Chamber of Commerce in Qatar carrying the same PANDEXIT and WISH story. Company-level coverage, no individual names. - [World Innovation Summit for Health](https://wish.org.qa): the awarding body, part of Qatar Foundation. - [LinkedIn recommendations, official record](https://www.linkedin.com/in/hec-ovi/details/recommendations/): 13 recommendations received, 6 from direct managers. LinkedIn issues no public per-recommendation URL and its Recommendation API is partner-restricted (as of August 2026), so the official record is this profile section, where every full text appears under its author's verified LinkedIn identity. Requires LinkedIn sign-in; LinkedIn serves HTTP 999 to crawlers, so a signed-in session is the only way to read it. - Five are quoted verbatim in the site's [Recognition section](https://hec-ovi.dev/#recognition). Their exact LinkedIn metadata, captured from the record above on 2026-07-28: - Christophe Billiottet, CEO and founder of ADGS, managed Hector directly. Written February 2024. Author profile: [linkedin.com/in/christophe-billiottet-25b415](https://www.linkedin.com/in/christophe-billiottet-25b415/) - Salvino A. Salvaggio, board-level advisor to ADGS, ex McKinsey. Written November 2024. Author profile: [linkedin.com/in/salvinosalvaggio](https://www.linkedin.com/in/salvinosalvaggio/) - Mark Sullivan, Product at Ohara, managed Hector directly. Written March 2026. Author profile: [linkedin.com/in/markjasonsullivan](https://www.linkedin.com/in/markjasonsullivan/) - Noah Madden, Founder of Seeker, managed Hector at Ohara. Written July 2025. Author profile: [linkedin.com/in/noahmadden](https://www.linkedin.com/in/noahmadden/) - Gonzalo Pedreira, Co-Founder of Contexto.ar, managed Hector at Dedo Interactive. Written November 2024. Author profile: [linkedin.com/in/gonzalopedreira](https://www.linkedin.com/in/gonzalopedreira/) ## Production work, live - [elcensuradoweb.com](https://elcensuradoweb.com): a Spanish-language AI news portal that publishes itself, in production and publishing daily (technology, politics, world, mystery, culture). An agentic brain with 32 MCP tools walks a 14-step gated editorial sequence on a 24/7 loop and falls back across agents (Gemini, Codex, Claude, a local model) when one fails. Synthetic authors carry their own bias and source lists. Append-only publish API, static reading path on a CDN. The site states its own AI-assisted nature in its editorial disclaimer. - [censurado-web-brain](https://github.com/hec-ovi/censurado-web-brain): Python control layer plus the 32-tool MCP server. - [censurado-web-backend](https://github.com/hec-ovi/censurado-web-backend): Go data and API server, sqlite store, append-only publish endpoint, operator panel. - [censurado-web](https://github.com/hec-ovi/censurado-web): Go static generator rendering the reading path. ## gamentic, the hackathon winner - [gamentic](https://github.com/hec-ovi/gamentic): browser-based AI dungeon RPG where the engine is the multi-agent system. A narrator orchestrates the world through 40 validated tools, every character runs as its own agent with isolated context and evolving memory, and the SQLite world state changes only through those tools, so long adventures stay consistent. Text on llama.cpp/Vulkan, images on FLUX.2 klein via ComfyUI, voice on Maya1 with per-scene emotion. Runs fully local on one AMD Strix Halo APU or cloud-driven. - [gamentic-anna](https://github.com/hec-ovi/gamentic-anna): the Anna-native port. The FastAPI engine and vanilla-JS UI stay intact, only the communication layer swaps to an Anna Executa with iframe transport. - [gamentic docs](https://hec-ovi.github.io/gamentic): "Docs and Blueprints", built from the live codebase. ## Agent tooling, published - [research-skill](https://github.com/hec-ovi/research-skill): project-scoped knowledge base that survives context compaction, with a contrarian pass that argues against the draft before it lands. On DeepResearch Bench II it lifts a Haiku 4.5 agent from 56.2% to 64.2% weighted, same model, same judge. - [websearch-skill](https://github.com/hec-ovi/websearch-skill): keyless multi-engine search and page reading. 9 keyless engines plus roughly 280 through self-hosted SearXNG, de-correlated rank fusion, Tor for .onion reading. Pages come back as paginated Markdown fenced as untrusted, so a page cannot steer the agent. - [websearch-skill on PyPI](https://pypi.org/project/websearch-skill/): published package, version 0.4.0. - [MCP Registry entry](https://registry.modelcontextprotocol.io/v0/servers?search=hec-ovi): io.github.hec-ovi/web-search, status active, stdio transport, uvx runtime. - [text-to-3D-skill](https://github.com/hec-ovi/text-to-3D-skill): a prompt to a rigged, game-ready GLB, fully local. FLUX.2 klein draws the concept, TRELLIS.2 builds the mesh on a Vulkan-only engine, headless Blender rigs a measured skeleton with idle, walk, run, jump. A face-count target keeps every output inside a game budget. 143 tests. - [agentickit](https://github.com/hec-ovi/agentickit): React copilot framework on AG-UI. The agent fills forms, opens dialogs, renders components, reads app state, calls your tools, and confirms destructive actions. A registry maps a different assistant per page or permission level. 579 tests, 4 runtime dependencies. - [@hec-ovi/agentickit on npm](https://www.npmjs.com/package/@hec-ovi/agentickit): published package, version 0.4.1. - [telegram-bot-skill](https://github.com/hec-ovi/telegram-bot-skill): any local CLI coding agent (Claude Code, opencode, Codex, Gemini) as a private Telegram bot. Reachable over MCP on stdio or Streamable HTTP. A deterministic tier gate (owner, trusted, guest, blocked) runs before any model sees a byte, onboarding is one QR scan. Zero-dependency Node core, 87 tests. - [blockchain-skill](https://github.com/hec-ovi/blockchain-skill): non-custodial wallet operations for agents on EVM and Bitcoin. Send, swap, bridge, sign, deploy and verify Solidity. Every contract is proven on a local EVM before a real network sees it, keys never leave an encrypted local keystore. - [ocr-skill](https://github.com/hec-ovi/ocr-skill): local image and PDF to Markdown OCR. A portable SKILL.md plus a self-contained CLI on DeepSeek-OCR-2 over llama.cpp, paginated for context windows, output fenced as untrusted. - [gmail-skill](https://github.com/hec-ovi/gmail-skill): Gmail behind plain agent commands. OAuth login, search, thread reading, drafts, send with attachments, labels. ## Private AI stack, local inference Measured on an AMD Strix Halo box (Ryzen AI Max+ 395, gfx1151, 128 GB unified memory). Every server exposes OpenAI-compatible endpoints. Benchmarks are published with their configs. - [vllm-awq4-qwen](https://github.com/hec-ovi/vllm-awq4-qwen): Qwen 3.6-27B AWQ-INT4 with DFlash speculative decoding, 24.8 t/s single stream, 256K context, vision and tool calling. Needed a custom HIP prefill kernel and 18 patches carried against vLLM. - [llama-vulkan-strix](https://github.com/hec-ovi/llama-vulkan-strix): llama.cpp servers on Vulkan for gfx1151. GGUF weights pinned to GTT let 128 GB of unified memory serve models past the VRAM budget. The opt-in ROCm FP4 plus MTP stack reaches 119 t/s decode on a 35B MoE, and a 118B MoE serves at 22 t/s. - [vllm-qwen](https://github.com/hec-ovi/vllm-qwen): BF16 OpenAI-compatible serving for the same board, 256K context on TheRock ROCm. - [rag-base](https://github.com/hec-ovi/rag-base): self-hosted hybrid retrieval with no LangChain. pgvector plus ParadeDB BM25 plus a LightRAG graph on Memgraph, BGE-M3 embeddings, four rerank modes, GLiNER NER, 19 endpoints, 115 integration tests. The published eval reports hit@1, hit@5 and MRR, including the configurations where the reranker makes results worse. - [comfyui-strix-docker](https://github.com/hec-ovi/comfyui-strix-docker): ComfyUI pinned to TheRock ROCm wheels on gfx1151, fixing the silent CPU fallback that stock Debian and Python 3.13 images hit. Base image for the generative projects here. ## Agentic workflows and loops - [open-research](https://github.com/hec-ovi/open-research): deep research as a pipeline of five agents passing work down the chain (Planner, Finder, Summarizer, Reviewer, Writer), on local inference, with SSE telemetry, durable sessions, PDF and Markdown exports. - [noob-cli](https://github.com/hec-ovi/noob-cli): compact Rust coding agent for local OpenAI-compatible models with a GPU front end. Concurrent sub-agents in their own tabs, plan mode, dockable windows, hardware monitor, every tool call inspectable, kernel folder lock via Landlock, live context-budget accounting. System prompt plus 14 tool schemas held under 2,000 tokens by a budget test. One 4.3 MiB static binary, 1,700+ tests. - [rebel-forge](https://github.com/hec-ovi/rebel-forge): self-hosted agentic social product. An 11-tool loop with error recovery runs research to draft to review to publish unattended across X, LinkedIn, Facebook and Threads, and reads engagement back the next cycle. LLM and image layers point at local endpoints or hosted vendors, swappable per layer. ## Generative and interactive - [ai-music-studio](https://github.com/hec-ovi/ai-music-studio): local album generation. An LLM plans the album, ACE-Step 1.5 generates the tracks, FLUX renders the cover art. SSE-streamed FastAPI backend, React frontend. - [duplexity-3d](https://github.com/hec-ovi/duplexity-3d): LLM-driven 3D adventure engine, a walkable three.js frontend over a local-AI backend. - [xubamp](https://github.com/hec-ovi/xubamp): classic Winamp 2.9x player rebuilt from scratch in Rust as a native Wayland client, no toolkit, PipeWire audio, classic .wsz skins. ## Engagements Four ways he is hired, each listed with its proof repositories in the [engagements section](https://hec-ovi.dev/#engagements). - Agentic Systems: agentic loops and workflows. Planning, tool calls, context engineering, output verification, recovery when a step fails, plus the skills and custom tooling the loop runs on. A workflow fits into a product you already have, a loop can be the product itself. - Agentic Engineering: 23+ years of software engineering, now orchestrating coding agents across any stack. Agents type, architecture and specs and integration tests and security review and deployment stay human. - Skills, Tooling and MCP: installable tooling for any agent, harness or CLI. Docker-isolated skills, npm and Python packages, MCP clients and servers over stdio or Streamable HTTP, from generative AI to web3. More than 10 tools published on GitHub, PyPI, npm and the MCP Registry. - Private AI Stack: inference tuned for your hardware plus the agentic harness to run agents on premises. Quantization, speculative decoding, custom kernels, context engineering, OpenAI-compatible endpoints so your code keeps working and your data stays put. ## Experience - Ohara, Senior AI Solutions Architect, contract remote, Mar 2025 to Jan 2026. Early vibe-coding platform, a hosted agent that turned a prompt into a deployed application. He replaced free-written API code with deterministic interface injection the model cannot deviate from, split the monolithic agent into a multi-agent system, wired MCP integrations across HTTP, SSE and WebSocket, wrote 30+ skill instruction sets before Agent Skills had a name, put ten multimodal vendors behind one validated interface, and built the web3 workflow that took agent-written Solidity through compile, audit, sign and deploy. - ADGS, Senior Full-Stack Developer, Doha, Qatar, Nov 2018 to Jun 2024. AI enterprise software for banks and government. Three products in production, each with a public product page: [PANDEXIT](https://www.adgs.com/pandexit), the agent-based epidemic simulator whose dashboard he built end to end (Leaflet map and heatmap streaming live state over WebSocket, a JSON-configured form engine driving the parameters); [STROKK](https://www.adgs.com/strokk), keystroke-dynamics behavioral biometrics, where he built the mobile capture end to end with a custom keyboard logging per-key timing and pressure plus motion sensors, one model trained per user; and [CARP](https://www.adgs.com/carp), capital assets replacement prediction, where he built the presentation layer. Also early OpenAI API integrations and hand-rolled tool calling before LangChain existed, and a dockerized NLP service comparing BERT, spaCy, regex and word2number head to head. - [ADGS](https://www.adgs.com): the company, a Qatar deep-technology firm working on agent-based simulation, cybersecurity and biometrics. - Grupo Ayex, Full-Stack and Game Developer, CABA Argentina, Nov 2015 to Nov 2018. Two certified casino games playing real money on dual-monitor slant-top hardware, device integration and C# services and ticket printing, a production vanilla-JS mobile roulette with OBS live video, and Azure Face API for age and identity verification at casino entry in production shape in 2015. - 2003 to 2015: games and interactive products. Globant for Electronic Arts on NHL and UFC (Lua and ActionScript, console certification constraints), a modular AS3 slot framework of his own design at Santo Mate with 4 titles in production, campaign and native Android work at Dedo Interactive, and Live Professor, a live video-teaching platform built end to end (AS3 plus Red5 streaming, PHP API, whiteboard, chat). Full timeline with the detail per role: [Experience section](https://hec-ovi.dev/#work). ## Stack - Languages: Python, TypeScript, Rust, Go, JavaScript. Earlier: C#, Lua, ActionScript, PHP. - Frameworks and infrastructure: React, Next.js, FastAPI, Node.js, Postgres, SQLite, Docker, Three.js, Phaser, Terraform with AWS. - Agentic systems: multi-agent orchestration and workflows, MCP servers and clients, tool schemas, context and prompt engineering, context budgets. - Applied AI: RAG, knowledge graphs, computer vision, multimodal generative AI across video, image, speech, audio and 3D. - Infrastructure: vLLM, llama.cpp, ROCm, Vulkan, CUDA, quantization and serving tuned per board. ## Formation - [IU International University of Applied Sciences](https://www.iu.org): Applied Artificial Intelligence coursework, 2023 to 2026, remote from Argentina. First year completed, 8 modules. Coursework, not a completed degree. Every module project is a public repo. - [snakeless](https://github.com/hec-ovi/snakeless): the Cloud Programming module. Serverless on AWS behind Terraform, Lambda, DynamoDB, S3 and CloudFront. - [healthhub](https://github.com/hec-ovi/healthhub): the object-oriented Python module. A habit-tracking application, Flask backend, Node frontend, MongoDB, Docker. - [computer-vision](https://github.com/hec-ovi/computer-vision): the Computer Vision module. YOLOv8, EfficientDet, Detectron2 and SAM2 benchmarked head to head for detection and segmentation in video, with a frame-by-frame analysis frontend. - Technical high school diploma in Personal and Professional Computing, EET N 468, Argentina. The rest is self-taught in production: Azure Face API in 2015, NLP before the hype in 2018, OpenAI APIs in 2022, local inference and custom kernels in 2024, agentic engineering as the delivery model in 2025. ## Availability and how he works - Available for remote work worldwide, async-first, from Rosario, Argentina (timezone America/Argentina/Cordoba). US, EU and AsiaPac overlap for sync work is normal, and he has been paid through EOR and contractor arrangements for years. - Contract is the default: defined scope, defined deliverable, defined window. Full-time is open for Applied AI Engineer or Forward Deployed Engineer roles at growth-stage AI companies, and he is open to CTO or technical co-founder for a team shipping a real product. - Rates: USD 90 an hour for contract work, USD 9,000 a month for full-time. Both cover architecture, implementation, review, and the agent orchestration behind delivery. - Not a fit: ML research or academic roles centered on model training, multi-week interview loops gated on live coding, teams that ban agentic engineering. - No live coding interviews: live coding measures hand-writing code instead of agentic engineering, which is how he delivers. Instead: an architecture or problem-solving discussion on a video call, and an architecture walkthrough of any project listed here. - Take-home assignments are consulting work, billed at his hourly rate of USD 90; the call and the architecture walkthrough are the free alternatives. - Contact through [LinkedIn](https://linkedin.com/in/hec-ovi) or the [contact section](https://hec-ovi.dev/#contact) of the site. Direct reply within 24 to 48 hours. ## Site sections One page, these anchors. - [Experience](https://hec-ovi.dev/#work): the three detailed roles plus the 2003 to 2015 timeline. - [Projects](https://hec-ovi.dev/#projects): featured work, then the full index by group. - [Agentic Systems engagements](https://hec-ovi.dev/#engagements): the four ways he is hired, each with proof links. - [Agent tooling](https://hec-ovi.dev/#repos): the published skills, packages and MCP servers. - [Private AI stack](https://hec-ovi.dev/#private-stack): local inference and retrieval infrastructure. - [Stack](https://hec-ovi.dev/#stack): languages, frameworks, capability groups. - [Recognition](https://hec-ovi.dev/#recognition): awards with their sources, and the verbatim LinkedIn recommendations. - [FAQ](https://hec-ovi.dev/#faq): 22 questions, grouped into the practice, capabilities, and working together. - [Contact](https://hec-ovi.dev/#contact): how to reach him. ## Machine files - [robots.txt](https://hec-ovi.dev/robots.txt): allow all, points at the sitemap. - [sitemap.xml](https://hec-ovi.dev/sitemap.xml): the single canonical URL plus its public images. - [og-image.png](https://hec-ovi.dev/og-image.png): 1200x630 social card. - JSON-LD in the page head: one schema.org @graph with WebSite, Person, ProfessionalService (the 4 engagements as an OfferCatalog) and FAQPage (all 22 questions).