Jaime Aza

A team of financial agents you talk to via chat — from WhatsApp, Telegram, Gmail or the web. Its promise: "it knows how you live, it invests like you are." Four specialized agents coordinate: (1) Expenses & income records and classifies transactions, generates monthly summaries and detects historical patterns; (2) Investments prepares stock operations on Wallbit but never executes them without your approval; (3) Market analyst contextualizes prices, trends and asset fundamentals according to your profile; (4) Risk profile measures your real tolerance from your history (not questionnaires) and blocks incongruent operations. You log expenses the way you talk — text, audio or a photo of the receipt — and Gmail sync captures transactions without manual input; search is in natural language, with no filters or exact categories, over embeddings in pgvector. The money lives in Wallbit (not in Tresqu): from the same chat you buy and sell stocks, ETFs and bonds, transfer and check balances live, with multi-currency support (COP, USD, EUR and more) and real-time conversions.

My own AI video, image and audio studio on top of the Higgsfield and ElevenLabs APIs, built agents-first: a REST API, an MCP server so Claude Code, Codex, Claude Desktop and ChatGPT can generate for me, and a web UI in Spanish (English optional) to do it by hand, all sharing the same library. It covers 82 Higgsfield endpoints (66 video, 15 image and 1 references), each with the JSON Schema extracted from the official docs, so the UI forms are generated from the schema with no per-model code and agents search by capability (text to video, image to video, first and last frame, references, editing, extension, motion transfer) rather than by brand. Audio comes from ElevenLabs: voice change on an exact segment of a video with horror effects, text to speech, sound effects, music and voice isolation, plus a sound library that classifies every sound so it can be reused without paying again. The core rule is that nothing is generated blind: the cost is shown next to the button and, over MCP, generating requires a single-use quote tied to that request and to an idempotency key, because the Higgsfield API does not deduplicate and an ambiguous retry could charge twice. Underneath are asynchronous, persistent jobs: validation before spending credits, a local queue sized to the account's concurrency (Higgsfield answers 400 rather than 429 at the limit), webhooks that only trigger an authenticated status check because they are not signed, and local copies of the outputs because Higgsfield keeps them for only 7 days. It includes presets with variables, batches with a total quote, a natural-language model recommender and 12 use cases with tutorials. Every iteration went through rounds of review by Codex before closing, and today it produces Frostbyte's reels alongside Blender, HyperFrames and DaVinci Resolve.

An infinite canvas for everything left pending while I work with AI agents. When Claude Code, Codex or another agent works on a task, more tasks keep coming out of it: a bug spotted along the way, an idea for later, a check before shipping. You stay on the main task and, by the end of the session, those side tasks are buried in the conversation. Parkboard collects them: when I say "let's leave that for later", the agent parks them with the `park` CLI, with enough context to pick them up cold and with the agent, session id and working folder they came from, so getting back to the conversation where they were born is one step. On the canvas, cards are grouped by project and moved by dragging, with status, priority, kind, notes, links, an Obsidian vault note and a short number (#12) so I can tell any agent "work on 12". Three views (Tasks, Ideas and Notes) keep ideas from crowding the work, plus personal and work areas. The CLI has no dependencies, returns JSON when piped, offers `--dry-run` on writes, and `park schema` describes every command, flag and exit code so an agent can use it without guessing. It is private by default: Clerk sign-in plus an email allowlist, and CLI keys are stored as SHA-256 hashes. Writes are chained per card and per project, with 409 conflicts, so the agent and I can edit at the same time without overwriting each other.

Proxy interceptor for Claude Code that closes the organizational alignment gap: Claude Code is aligned with Anthropic's values, but not with the policies of each company deploying it. Tranquera is the middle layer — a silent customs checkpoint that enforces no-code rules at runtime with a Regex → Pattern → Haiku 4.5 judge cascade under 200ms of overhead and four explicit actions (BLOCK · REDACT · WARN · LOG). The (non-technical) compliance officer defines policies in natural language from a visual builder; devs onboard with a single command (`npx tranquera setup`, Google OAuth device flow and automatic export of `ANTHROPIC_BASE_URL` to the shell rc). Per-dev attribution is solved by baking the token into the URL path — Claude Code does not allow custom headers. An AI Suggestor with a daily cron and Haiku 4.5 proposes new rules from the LOGs with a human in the loop. Postgres + pgvector ready to pre-filter rules by semantic similarity. Winner of the AI Security track and the overall prize at Platanus Hack 26 Buenos Aires (main sponsor: Anthropic).

Professional gallery delivery platform for photographers with state-of-the-art multimodal AI. It automatically analyzes every photo with GPT-4o and Gemini 2.0 Flash: assigns a 1-10 score, evaluates composition, pose and background quality, categorizes scenes (ceremony, portraits, party, outdoors...) and generates project highlights. It includes semantic photo search in natural language ('bride dancing', 'table decoration') with multimodal RAG using Google's Gemini Embedding 2 — Google's first natively multimodal model mapping text and image into the same vector space — stored in pgvector. The public gallery reorders photos based on the client's favorites and writes poetic cover phrases with AI. Complemented by festora-vision-api: an independent microservice with blur detection (Laplacian), BRISQUE/NIMA technical scoring, emotion analysis (DeepFace) and CLIP/DBSCAN semantic clustering.

Presential-event networking that finally becomes visible: you scan someone's personal QR and end up inside the event and connected in the same gesture — no 'add' buttons, no requests, no connecting from the couch. Every edge in the graph represents a real encounter (it's only born by opening someone's QR, never on render), and the app spreads person to person through contagious membership: getting into an event always goes through someone who's already in. The event's full constellation — visible to every attendee, not 'my network up to 2nd degree' — is drawn live with react-force-graph-2d on Canvas, synced in real time via Supabase Realtime, with one star per attendee and triadic closure when three people who already know each other form a triangle. Auth is Google OAuth only, role/interest/intent tags are a living catalog that grows with usage, and onboarding is a single screen designed to be used standing up, one-handed, amid event noise. Built with Next.js 16 (App Router, React 19, React Compiler) and Supabase (Postgres + RLS, Auth, Realtime).

The full operating system of a bar and restaurant, with AI at the core of the product rather than bolted on: it takes orders, sells, writes the menu and decides what can be ordered. Orders come in through three doors that end in the same kitchen ticket: the QR code on every table, the customer app and an AI agent that answers the business's WhatsApp around the clock. That agent is built with LangChain create_agent on LangGraph and GPT-4o-mini, with tools wiring it into the business: browsing the menu and each product, searching regardless of accents or typos, reading the customer's history and preferences, quoting, checking delivery coverage, creating, modifying and cancelling orders, and asking for a human. Its conversational memory lives in PostgreSQL through PostgresSaver (one thread per contact per day), while long-term context (address, preferences, past orders) comes in through tools. It understands voice notes with gpt-4o-mini-transcribe and photos with multimodal vision, payment receipts included, pulling out amount, date and reference number. Before creating a delivery it validates the address against the coverage area hand-drawn on the map (ray casting in Django, Turf on the front end) and, when the conversation goes off script, it hands off to a person and pauses without losing the thread. The rest of the product is AI too: a professional product photo generator with four swappable models (GPT Image 1.5 and 2, Gemini 3 Pro and 3.1 Flash Image) on Cloudflare R2, AI-written product descriptions, a menu recommender driven by mood or a quick quiz, voice search with Whisper and daily content that refreshes itself. Underneath runs the real operation: live kitchen over WebSocket (Django Channels), tables per floor with printable QR codes, inventory with recipes and real cost per variant, expenses that separate investment from operating spend so the margin doesn't lie, analytics, reservations, music per floor, multiplayer games and two businesses in the same account, with no per-order commissions. Running in production every day in a real venue in Cumbal (Nariño) and the base of the SaaS being packaged for other businesses.

Real-time multiplayer typing race inspired by Monkeytype. Anyone can create a room with a 5-character code, share the link and compete for the highest WPM on the same text, with a spectator mode to join rooms mid-race. The architecture runs 100% on the edge over Cloudflare Durable Objects: one 'Room' DO per room (authoritative state, WebSockets with the Hibernation API, persistent hostToken) and a 'RoomRegistry' singleton that keeps the global list of active rooms via RPC between DOs. The clock, the text and the ranking are decided server-side, so cheating by changing the local clock is impossible. Includes 5 themes, key sounds synthesized with the Web Audio API (no samples) and a host/player/spectator model with kick and race restart.

Windows desktop app that transcribes speech to text in real time and types the result directly wherever the cursor is — no copying, no pasting. Push-to-talk with F9: hold to speak and release so the corrected final text appears in any app (VS Code, Word, Notion, Chrome…). Supports two swappable engines: Deepgram Nova-3 (cloud, low latency, simultaneous es+en bilingual with LANGUAGE=multi) or local faster-whisper (no internet, no API key, tiny→large-v3 models). Includes a Vercel-style floating HUD with timer, live preview, voice visualizer and similarity-based deduplication. Lives as a system tray icon, with VAD via silero-vad ONNX (no PyTorch). Packaged with PyInstaller + Inno Setup as a native .exe installer.

Web app that converts a 2D image into a navigable 3D point cloud using AI depth estimation. The backend (FastAPI + Docker on Railway) wraps Depth-Anything-V2 (ViT-S checkpoint) and returns the depth map as a grayscale PNG. The frontend combines the RGB image and the depth map in utils/pointCloudUtils.ts: for each pixel it generates a 3D point where X,Y map to normalized space (NDC-like, with inverted Y and height/width aspect ratio) and Z = (depth/255) × depthScale, controlled by sample rate and depth exaggeration sliders. The result is two Float32Arrays (positions + colors) rendered with React Three Fiber + drei as a THREE.Points with vertexColors. It supports three swappable depth backends: a self-hosted FastAPI server, HuggingFace Spaces or the OpenAI Images API.

Real estate semantic search engine with a RAG architecture covering more than 8,800 properties in Mexico. The user asks in natural language ('spacious house with a garden in Mérida, 3 bedrooms'), Gemini 3 Flash extracts structured filters (city, type, bedrooms, price), Qdrant pre-filters and vector search retrieves the most semantically relevant properties. It supports multimodal RAG with Google's Gemini Embedding 2: each property is represented by a single 3072-dimension vector fusing the text description with up to 6 images in the same vector space — enabling search by text or by image. It also supports swappable embeddings (OpenAI text-embedding-3 and gemini-embedding-001) with a React playground featuring an interactive similarity chart and a conversational agent with LangGraph (ReAct). Deployed on Railway + Vercel + Qdrant Cloud.