ROADMAP.md

Personal LLM - Roadmap

Phase	Scope	Key deliverables	Success metric
v0.1 Spine (done)	Memory + RAG + KG-lite + CLI/API/Streamlit + tests/CI	Model router, SQLite memory store, Chroma ingest/retrieve, grounded RAG with citations, minimal KG, 3 interfaces	Ingest real notes; ask questions; get cited answers; offline-capable for ingest/retrieve
v0.2 Agent + Tools (done)	Base Agent loop, typed tool layer, 4 safe tools across all 3 permission tiers, permission + audit	Agent class (plan-act-reflect, structured JSON steps), ToolRegistry, memory_search/remember/read_file/web_fetch tools, audit log via store.log/store.recent_audit	Agent completes a multi-step task using tools, fully logged; every tool call permission-gated; unauthorized/invalid calls fail as an observation, not a crash
v0.3 Proactive + Router+ (done)	Proactive review job, richer router (multi-provider verification)	review/weekly.py (recent + forgotten-but-important surfacing, stored as new memory), ModelRouter.complete_with_verification() (agreement via local embeddings, no paid provider needed), ask --verify	Agent/review surfaces something useful unprompted; disagreement between providers is flagged, not silently picked
v1.0 Personal assistant (done)	Real integrations (Gmail/Drive via existing MCP), installable package, daily-use polish	integrations/ (credential-free ingestion, idempotent by doc_id), cli ingest-external, POST /integrations/sync, [project.scripts] console entry point (personal-llm), --version, doctor (provider/data health check)	Real inbox/drive-aware answers proven live with real Gmail threads + a real Drive doc; pip install -e . produces a working personal-llm.exe; doctor is the first command to run before daily use
v2 Voice + Vision (done)	Local STT/TTS, OCR image ingestion, Gemini vision Q&A - reclassified from speculative to built (ADR 0006)	voice/stt.py (faster-whisper), voice/tts.py (pyttsx3), vision/ocr.py (pytesseract, Tesseract-optional like Ollama), ModelRouter.describe_image(), cli transcribe/ask-voice/describe-image, POST /voice/transcribe, /voice/ask, /vision/describe, /vision/ingest	TTS->STT round trip verified live with real local models (no mocks); OCR/Gemini-vision verified live too, correctly degrading since Tesseract/API key aren't provisioned yet
v3 (speculative)	Multi-device sync, SDK/plugin marketplace, continuous learning loops, voice cloning, live multi-speaker	Architecture allows it; not designed in detail yet	Revisit only if there's a second user/device, or a specific need for the harder voice/vision capabilities

Milestones & effort (v0.1 -> v1.0)

Milestone	Engineering effort	Primary risk
Router + memory store	Small (1-2 sessions)	Python 3.14 vs 3.12 wheel mismatch
Ingest + retrieve + RAG	Medium	Retrieval quality on small corpora
KG-lite	Small	LLM extraction noise/cost
Interfaces (CLI/API/Streamlit)	Small-Medium	None significant - proven pattern
Tests + CI	Small	Mocking router cleanly
Agent + tools (v0.2)	Medium-Large	Tool safety, permission design
Real integrations (v1.0)	Medium	Auth/OAuth complexity, rate limits
Voice + vision (v2)	Medium	External Tesseract binary not pip-installable; vision Q&A is Gemini-only

Risk Assessment

Risk	Likelihood	Impact	Mitigation
Python 3.14 breaks ML wheels (chromadb/torch)	High if ignored	Blocks build	Dedicated 3.12 venv, documented in CLAUDE.md
Scope explosion (JARVIS trap)	High	Nothing ships	Hard P0-P3 triage in PRD; v0.1 = memory+RAG only
Ollama not installed / no GPU	Medium	Local path unavailable	Optional dependency; Gemini fallback always available
Free-tier rate limits (429s)	Medium	Flaky demos	Retry/backoff reused from analyzer.py
Prompt injection via ingested/web content	Medium	Bad tool actions (v0.2+)	Untrusted-content tagging from the first tool built
Privacy expectations vs implementation	Low-Medium	Trust erosion	Local-first by default, documented data flows, audit log
Solo-maintainer burnout on a huge vision	High	Project abandoned	Ship small, working slices every phase; v0.1 has a hard finish line
Tesseract OCR binary not installed	High (not installed by default on Windows)	Image ingestion unavailable	Optional dependency exactly like Ollama; clear actionable VisionError instead of a crash

Future Evolution

5-year vision: Personal LLM is the substrate every one of Zaid's AI projects imports; it has years of consolidated memory, a rich personal knowledge graph, a handful of trustworthy agents (research, coding, career) operating under clear permissions, and optional voice/vision modes for hands-free use.

10-year vision (speculative, not designed): A fully proactive digital twin capable of representing the user in low-stakes interactions (scheduling, first-pass email triage) with strong auditability, still local-first and user-owned - explicitly not a chase of "replace the user," but an amplifier under their control.