ProfMe

An AI-native professional intelligence product owned end to end — customer problem, constrained feature design, durable extraction pipelines, human review boundaries, and post-ship iteration.

Evidence → Memory → Intelligence → Action

01 — Overview

ProfMe is an AI-powered professional intelligence system. It acts as a unified layer that observes professional activity (meetings, calendar, communications) and transforms it into structured, actionable context.

02 — Problem

Professionals spend hours synthesizing information across meetings, chat, and email to determine what was decided and what needs to be done. Traditional AI tools simply summarize transcripts or provide a chat interface, failing to maintain a persistent professional state that is continuously updated from multiple evidence sources. The real challenge is preventing AI-generated assumptions from silently becoming truth.

03 — Why it was technically difficult

Building a reliable professional intelligence system requires solving several hard problems:

• State management: Maintaining a canonical source of truth that is incrementally updated by probabilistic models.
• Entity resolution: Resolving people, projects, and context across fragmented data streams.
• Provenance: Ensuring every AI-derived commitment or decision can be traced back to the exact moment in a meeting or conversation.
• Workflow coordination: Orchestrating complex, long-running extraction pipelines that interact with external APIs, audio processing, and LLMs without losing state on failures.

04 — Architecture

Professional activity (meetings / calendar / communication / work systems) → Evidence → Structured professional memory → Decisions / commitments / tasks / open loops / people / projects → AI reasoning → Today / Review / actionable context

The system is built on a rigid relational database (PostgreSQL) that acts as the single source of truth, synchronized with local desktop clients.

05 — Key engineering decisions

• Evidence-backed AI: We separated raw evidence from derived intelligence. Models never overwrite evidence; they only propose structured intelligence linked to the source.
• Temporal workflows: Implemented Temporal to manage asynchronous extraction pipelines, ensuring idempotency and reliable failure recovery.
• LLM Gateway Architecture: Used LiteLLM to abstract provider APIs, allowing dynamic routing and fallbacks for critical extraction tasks.
• Human-in-the-loop boundaries: Built explicit review workflows where AI-generated tasks and decisions remain in a pending state until approved by a human.
• Strict Data Modeling: Used PostgreSQL to enforce schema constraints on AI outputs, rejecting malformed structured data before it enters the system.

06 — AI architecture

Input (Audio/Text) → Speech Gateway → Structured Extraction (LLM) → Validation (Pydantic/Postgres) → Professional Memory → Human Review → Actionable Task

07 — Product decisions

• Review over automation: We chose to make the system propose prepared actions rather than taking actions automatically on the user's behalf.
• Transparency: The interface visually distinguishes between raw facts, derived intelligence, and AI-generated suggestions to build user trust.
• Desktop synchronization: Implemented local-first sync to ensure the tool is lightning fast and available when offline or in environments with poor connectivity.

08 — Outcome

ProfMe successfully processes complex meeting audio and work activity, generating high-fidelity structured professional context. The system maintains strict provenance, allowing users to trust the generated tasks and decisions.

09 — What I learned

I learned that building production AI is less about prompt engineering and more about distributed systems engineering. Ensuring reliability requires building robust boundaries around probabilistic models, handling failures gracefully, and enforcing strict data schemas.