Fiona: Custom Personal AI Companion

Alfija advises C-suite executives on AI transformation, and wanted to know, viscerally, what it costs to design a personal AI agent that belongs to a single human, not a vendor. Not a chat-window subscription. Not a generic assistant prompt-tuned with a name. An actual personal companion: persona engineered as a design artifact, memory persistent across days, voice in and voice out, web-grounded when it does not know something, running on infrastructure she owns. Tracking what she would do for a CIO building an internal personal-AI surface, applied to herself first. The point of Fiona is not to replace a phone or a search bar. The point is the discipline of an agent that is designed, file by file, and owned end to end.

What it is concretely. Fiona lives in Telegram. The user sends a message, typed, voice-noted, or photo, and Fiona responds in the same surface, in voice or text depending on the moment. The exchange feels like a conversation with someone who remembers what was said yesterday, because she does: a daily memory log accumulates per session and is indexed for semantic recall, so the next time the same topic comes up the agent does not start from zero. Behind the chat is a multi-agent runtime, a transcription pipeline for incoming audio, a Swedish-trained TTS for the reply, and a web-search loop the agent reaches for when the question wants current information. Underneath all of it, a single European server she runs herself.

Persona as engineering, not prompting. Most personal AI products ship a persona by adjusting one prompt. Fiona is composed across separate files that each carry a distinct concern, a soul file describes the inner character (helpful, curious, playful), an identity file defines what the agent is and what it is not, a user file holds what the agent should know about the person it works for, and a behavioural manual codifies how it should act when ambiguous. Each file is a design artifact under version control, edited deliberately, with the same care a product team would put into a brand-voice document. The persona is not act friendly. It is a structured object the agent loads on every turn and behaves consistently against. When the user wants the persona to drift, warmer, more terse, more direct, the change is made to a file, not to a session.

Memory as a first-class system. A daily memory log is written as the conversation flows, capturing the substance of what was discussed rather than the literal transcript. The logs accumulate, and a semantic-embedding layer indexes them so the agent can pull relevant past context when a new turn relates to something said weeks ago. The user can ask Fiona what she remembers about a topic and get an answer drawn from the index, not improvised, memory is queried before the reply is composed. Memory is not a chat-history scroll; it is an asset the agent maintains and consults.

Multimodal by default. Voice notes are transcribed and handled the same way as typed messages. Replies are spoken when speaking is the right modality and written when reading is. Photos can be sent and described. The user picks the modality of the moment, driving, cooking, late on the couch, and the agent matches it without being told to switch. The pipeline is wired so a single conversation can move between text and voice freely; modality is a property of the message, not the session.

Web-grounded when it does not know. When the user asks for something the agent does not have an answer to, current news, a fact worth verifying, a venue's opening hours, Fiona fetches it live through a web-search tool and answers from the result, not from improvisation. Old information labelled as old; current information labelled as current. The improvement over a stock chatbot is not theatrical; it is the difference between an answer that is right today and an answer that was right at training time.

Owned end to end. The agent runs on a single European server she controls. No third-party admin platform sits between her and her own conversations. No vendor logs the transcripts. Nothing in this surface feeds any company's training data. For a personal AI agent, one that knows her schedule, her habits, the names of her people, that ownership is not a marketing line; it is the architecture choice that makes the persona safe to develop in the first place.

Why this lives in the AI portfolio and not the personal-projects drawer. Because it is a working answer to the question CIOs ask her every other week: what does it look like when an executive has their own AI surface, not a vendor product, not a corporate seat, not a SaaS subscription, and what does it cost to build one? It looks like a Telegram thread with a persona that was composed file by file, a memory that grows on disk she owns, a voice pipeline that can be heard with closed eyes, and a web-search tool the agent reaches for when it does not know. The same agent runtime that runs the nine-agent Emmis Company stack and the Björn health intelligence system also runs Fiona, different design, different purpose, same discipline.

What you see when you message Fiona is a calm, curious voice that remembers you. What's underneath is a deliberately composed persona, a per-day memory log indexed for semantic recall, a multimodal in/out pipeline, a web-search grounding loop, and a server she owns the keys to.