We don't fine-tune. We compile.

Three parallel streams during the notice period:

• Interview transcripts. AI-led sessions are recorded and transcribed with Whisper. Per-session structured prompts hit the role's known gaps and pull edge cases.
• Passive observation. OAuth read-only access to email, Slack, calendar, and file storage. The Observer agent scans for recurring patterns — meetings, vendor exchanges, decision points — that the employee may not have mentioned.
• Artifact harvesting. Documents, spreadsheets, dashboards the employee owns or frequently touches.

Everything lands in a per-tenant capture lake on encrypted storage, scoped to the consented window.

A pipeline of focused LLM passes converts the capture lake into structured knowledge. Each pass has a narrow job:

• Process extraction: trigger → steps → exceptions → stakeholders → cadence.
• Decision rules: "when X, the employee did Y because Z."
• Stakeholder map: people, their roles, the rules of engagement, escalation paths.
• Vendor & tool registry: who you pay, what for, account contacts, undocumented terms.
• Quirks & tribal knowledge: things the policy doc says vs. what the employee actually does.
• Recent incidents: what burned them, how they fixed it, the lesson.

Each extraction cites its source (interview line, email thread, calendar event). The departing employee reviews every entry and can correct, redact, or annotate.

Output is a per-role structured document we call the Role KB — typically 30–80 pages of compiled knowledge. Stored as:

• Markdown / JSON for human readability and machine queries.
• Embeddings (pgvector, OpenAI text-embedding-3) for semantic search.
• Source citations bidirectionally linked back to the capture lake.

The Role KB is the thing you own. You can export it, audit it, share it with HR or the new hire's manager, or feed it into your own internal knowledge tools.

When the successor asks a question, we run a hybrid retrieval:

• Structured query on the Role KB for known process / vendor / stakeholder lookups.
• Vector search across embeddings for semantic / fuzzy questions.
• Re-ranking with a cross-encoder to surface the most relevant chunks.

Top-K chunks are injected into the successor agent's context with their citations. The model answers grounded in retrieved content — not from training memory.

A persistent chat interface backed by Claude. Its system prompt encodes:

• The departing employee's voice and register (calibrated during capture).
• Hard rules: cite sources, flag uncertainty, never invent processes.
• Escalation paths: low-confidence questions route to the standby Slack channel during the 30-day post-handover window.

Every successor message is logged. The transcript becomes ongoing fine-tuning data for the Role KB — not the model.

Before handover, we run a calibrated eval:

• 20–30 role-specific questions, written by the departing employee.
• Both the employee and the agent answer.
• The employee grades both blind on factual accuracy and judgment fit.
• Match rate target: ≥85%. Below threshold = iterate on extraction, not the model.

Post-launch, we track resolution rate (successor question → useful answer, no human escalation) and hallucination rate (incorrect or invented answers). Both are visible in the customer dashboard.

Five reasons we explicitly chose RAG over fine-tuning:

• Privacy. Per-tenant Role KBs mean no cross-customer data ever touches model weights. Fine-tuning creates that risk.
• Cost. A single Role KB compile costs $50–150 in LLM tokens. Fine-tuning per role would 10–50× that, with weeks of latency.
• Update latency. When a process changes, we patch the KB and it's live in seconds. Fine-tuning would require retraining.
• Honesty about uncertainty. RAG models can say "I don't have this in the KB" — fine-tuned models confabulate confidently. For business operations, confabulation is the failure mode that kills trust.
• Portability. Your Role KB is yours. Export it, take it to another tool, use it with a different model. Fine-tuned weights would lock you in.

Honest numbers for a single departing employee, 4-week capture, 12 months of successor use:

Gross margin at our published prices: ~85%. The economics scale linearly per seat, which is why BPOs at 50–500 seats per contract are our wedge.

Production (private beta): Interview transcripts, extraction pipeline, Role KB, successor chat, Turing eval, per-tenant isolation.
Q3 (private beta+): OAuth observation (Gmail, M365, Slack), file artifact harvest, customer dashboard with resolution / hallucination rates.
Q4: SOC 2 Type I, EU residency, multi-language, expanded eval surface.
2027: Continuous capture (Option A — full role twin from day 1), agentic successors that can execute bounded tasks, not just answer.