OpsPilot

01 / The Problem

I run finance and operations at a tech company. My day is split across Teams, SharePoint, Xero, HubSpot, and email. Someone asks a question in one system, the answer lives in another, and the context for why it matters is in a third. I was spending hours just pulling data, missing follow-ups because nobody tracked who said what, and scrambling for context before meetings.

"What does the Customer X contract say about payment terms?" That's a SharePoint search. "Who hasn't paid us?" That's Xero. "What deals is the sales team working on?" HubSpot. "Tell me about Customer X." That needs all three. The information always existed. It was just scattered across 6,000+ files, multiple systems, and years of accumulated documents.

Before

Dig through 6,000+ SharePoint files manually
Log into Xero, HubSpot, email, calendar separately for every question
Miss follow-ups because nobody tracked who promised what
Scramble for context 5 minutes before a meeting

After

Ask a question in Teams, get an answer from any source in seconds
Morning briefing with everything that matters before 9am
Commitment tracker catches every promise, nudges when overdue
Pre-meeting brief lands 15 minutes before every call

02 / How It Works

Nine-stage pipeline

Someone asks a question in Teams. The pipeline triages it, figures out whether it needs documents, live accounting data, CRM data, or all three at once, and returns an answer as an Adaptive Card with approve/edit/reject buttons. Live data questions hit Xero or HubSpot directly, zero LLM cost. Document questions run a hybrid search (BM25 + FAISS + reranker). Cross-source queries like "tell me about Customer X" pull from everything simultaneously. A PII blocker catches sensitive data before anything goes out.

03 / What It Does

What it does

Hybrid Search

BM25 for keywords, FAISS with E5-Large-Instruct for semantics, Reciprocal Rank Fusion merges rankings, then a ROSES AI reranker picks the top five. HyDE generates domain-specific terms so vague queries like "anything about the Customer X deal" still find the right documents. 100% accuracy on production tests. The search pipeline is frozen. No changes without explicit approval.

Cross-Source Intelligence

"Who hasn't paid us?" pulls live Xero receivables. "What deals is David working on?" queries HubSpot pipeline data. "Find the employment contract template" searches SharePoint. Ask "tell me about Customer X" and it pulls from all three simultaneously. 17 direct API actions, and when a query routes to live data the LLM never touches it. Just formatted numbers, straight back.

11-Page Dashboard

Built on McKinsey's Vizro framework. Covers the full picture: P&L, cash position, sales pipeline, receivables, CRM activity, engagement trends. 40+ chart types. There's an AI chat overlay too, so you can ask Claude to recut any chart without leaving the page.

Autonomous Monitoring

10+ scheduled jobs that run without any manual trigger. Morning and evening briefings land in Teams with calendar, email highlights, financial alerts, and follow-up reminders. Deal stage monitoring checks HubSpot every 5 minutes during business hours. An invoice scanner picks up PDFs from my inbox, uses Vision AI to extract line items, matches against Xero records, and flags duplicates or price drift. A commitment tracker detects promises in Teams and email, then alerts when they're overdue. Pre-meeting briefs generate 15 minutes before each call with attendee CRM history and open invoices.

Personal Growth Journal

Every evening, it writes a first-person diary entry from the day's activity: Teams conversations, emails, calendar, deal changes, financial updates. A separate behavioural feedback system analyses my communication patterns and flags blind spots like unreplied messages, tone issues, or neglected relationships. Rolling memory tracks per-person interaction metrics and recurring feedback themes. Weekly digest every Friday with trends and self-review evidence.

Conversation Memory & Approval

It remembers what you were talking about. Per-user context with a 2-hour window and 10-turn memory. Ask about Customer X, then follow up with "what about their payment terms?" and it knows what you mean. Every answer comes back with options and nothing goes out without you approving it. It also learns which style of answer you tend to pick from your approval history.

04 / In Action

Teams Adaptive Card

Ask a question in Teams, pick an answer, click approve. Nothing gets sent without you seeing it first.

Sample data

Microsoft Teams

OpsPilot | General

What's our current MRR and how does it compare to forecast?

OP

OpsPilot 10:42 AM

Monthly Recurring Revenue

MRR is £127,400 as of March 2026, 4.2% above the Q1 forecast of £122,300. Growth is driven by 3 new enterprise contracts closed in February. Churn held at 1.8%, below the 2.5% target.

Current MRR£127,400

Q1 Forecast£122,300

Variance+4.2% above target

Churn1.8% (target 2.5%)

Approve Edit Reject

05 / Built With

Tech stack

Python 3.12

59k+ lines across 173 files

Claude Sonnet

Reasoning, synthesis, orchestration

Claude Haiku

Triage, OCR, fast tasks

FAISS

Vector similarity search

BM25

Keyword search engine

MS Graph API

Teams, SharePoint, Calendar, Email

Xero API

Accounting data (read-only)

HubSpot API

CRM & sales pipeline

Bot Framework

Teams webhook handler

Vizro

11-page McKinsey dashboard

APScheduler

11 automated briefing jobs

Oracle Cloud

Ubuntu VM, systemd, auto-deploy

More Work