Autonomous agent that reads every onboarding signal — Slack, Jira, email, calls, docs, CRM — and produces evidence-cited briefs, calibrated slip-risk forecasts, and review-ready customer emails. Catches slips before they happen, with every claim grounded in a verifiable source.
A single Implementation Manager at a mid-market B2B SaaS company carries 8–15 concurrent onboardings, each running 30–180 days and throwing off evidence across Slack Connect channels, Jira projects, recorded weekly calls, email threads, kickoff decks, screenshots, and the CRM. Reconstructing the true state of each account — forecasting which will slip, writing an honest weekly customer update — burns 8–12 hours a week of manual synthesis per IM and still ships reports that are stale by Monday. Slip detection is gut-feel; blockers surface weeks after they started; and onboarding failure is the single largest driver of early-stage churn, so every missed signal is recurring revenue walking out the door. Incumbent tools (Rocketlane, GUIDEcx, Vitally) are project-management dashboards with summarization bolted on — not reasoning systems that read the evidence and stand behind every claim. This is the production-grade agent platform underneath the category.
One adapter protocol fuses Slack, Jira/Linear, Gmail, Calendar, Salesforce, recordings, screenshots, and kickoff PDFs. Audio is transcribed, images are OCR'd, documents are parsed — then everything normalizes to a single evidence schema, deduped twice (source-key constraint + content SHA-256) and embedded before persistence. The agent never branches on modality.
Every checkpoint fuses three retrievers — Postgres full-text search, pgvector cosine similarity, and a recency-ordered structured filter — via reciprocal-rank fusion. Pure-vector was explicitly rejected: lexical recall and recency matter as much as semantic similarity when the corpus is a live operational feed, not a static knowledge base.
A purpose-built orchestrator — retrieve → reason → cite → reflect → slip_score — with a pluggable step protocol and no black-box framework. It generates daily briefs, weekly status, or drafted customer emails; binds every claim back to the evidence rows that support it; reflects and drops anything it can't cite; then forecasts slip risk green / yellow / red with weighted reasoning. Every step's input and output is captured in the run trace.
No floating confidence scores. Every claim in every brief links to the exact Slack message, ticket, or meeting moment behind it. The operator console's workbench hover-couples each claim to its evidence — fully client-side, zero-network — so a reviewer verifies provenance in one glance before anything reaches a customer.
A first-match-wins YAML policy engine routes each output (auto-publish, route-to-review, escalate). Customer-facing drafts land in a human-in-the-loop queue, and every decision writes an immutable before/after audit row. A backtest harness replays the agent over historical checkpoints and scores slip calibration, blocker recall + lead time, citation precision, and brief acceptance — gating CI on regressions.
The whole platform is published over the Model Context Protocol, so Claude Desktop, an IDE, or another agent can drive it as six tenant-scoped tools — list/get implementations, hybrid evidence search, generate a brief, fetch a brief with its claims + citations, and read the latest slip band. Every call resolves and authorises against the named tenant, so no tool can read across tenants. Run it with one command: `iih mcp`.
LangGraph, Pydantic-AI, and the Anthropic Agent SDK were all evaluated and rejected in favour of a compact orchestrator behind one step protocol. The payoff is total transparency — every reasoning step is inspectable in the run trace — and zero lock-in. Enterprises auditing an agent for customer-facing work need to see exactly what it did and why.
Reimplements scheduling and retries the frameworks hand you for free. Because every step lives behind one interface, swapping in a framework later is a contained change, not a rewrite.
Every claim carries citations to the source records that ground it; the reflect step deletes any observation it can't back up. Reviewers jump to the exact span; downstream systems verify provenance. The platform earns trust by being checkable, not by asserting a number.
The cite/reflect passes cost extra model calls and add loop latency. Non-negotiable for anything that touches a customer.
Per-tenant credential keys are derived via HKDF-SHA256 over the platform key plus the tenant id — a ciphertext encrypted for one tenant provably cannot be decrypted under another's. Row-level isolation by tenant_id is on every table, and connector accounts are per-tenant. This is the security posture enterprise procurement actually checks.
One-way key derivation means rotating the master key invalidates stored credentials; a re-key flow is a clean future addition.
A transparently-heuristic mock LLM, deterministic embeddings, and a seeded synthetic dataset let the entire platform run end-to-end in under a minute with zero API keys or SaaS credentials. The instant a real key is configured, the identical loop runs against Claude and live connectors — the architecture doesn't change, only the providers behind the interface.
Heuristic output trails a frontier model until the key is set. The point is to lock the platform shape and let any evaluator reproduce it instantly.