Helping automate Investment Firm Workflows

# Senior Automation / AI Engineer — Build an Automated Daily Meeting Briefing System **Type:** Freelance contract · Remote · Milestone-based **Stack:** Python or Node, LLM APIs, Gmail + Google Calendar APIs, HubSpot API, Railway (or similar hosting) ## Who we are A small, lean investment firm (family office). Our principal's calendar is packed with external meetings — first calls with founders and fund managers, follow-up meetings, investor and portfolio company check-ins. Today a team member hand-builds a daily briefing document covering every meeting. We want that fully automated, at the same quality bar, and we have a detailed spec, worked examples, and calibrated prompts ready to hand over. ## What you'll build (V1) A nightly automated pipeline that does two things: **1. Closes out today** - Summarizes the day's call transcripts (from our meeting recorder) into post-call notes - Stages follow-ups and action items into HubSpot — approval queue at first, automatic once trusted - Reads the day's new email once and files each message against the right relationship, building a running chronological record per counterparty (this replaces re-searching mailboxes every night, which is where past attempts have failed) **2. Builds tomorrow's briefing** - Pulls the principal's calendar for the next day - Resolves who each attendee actually is across calendar emails, transcript speaker labels, HubSpot records, and email display names (these rarely agree) - Assembles one entry per meeting in our house format, templated by counterparty type and meeting stage: header, current deal/raise status from email context, a memo built or updated from their materials using our existing memo prompts, team backgrounds with LinkedIn links, recap of the previous call, who introduced them, and that day's action items - Runs a cached research pass on counterparties we've never met **Delivery** - Clean PDF plus markdown source, in the inbox by early morning (hard deadline midday), with a full searchable archive - Drafts go to a human reviewer first; the review window shrinks as the system earns trust - Corrections loop by reply email ("that's a different person") — fixed once, fixed permanently **Stretch item:** a draft "questions to dig into" section per meeting, grounded in relationship history. Delivered as draft-for-edit; dropped if it doesn't earn its place. **Explicitly out of scope for V1:** full multi-party meeting treatment (diarization limits), and follow-on modules (browsable relationship database, portfolio marks tracker, deal-grading rubrics) that will be scoped separately once V1 is live. ## Validation before go-live Nothing reaches the principal until it passes three gates: 1. Regenerate our hand-built example entries from the raw inputs (transcripts, decks, email chains) and match them side by side 2. Two-week backtest against real calendar, email, transcripts, and CRM data, graded against what we actually built 3. Shadow run in parallel for 1–2 weeks before cutover ## Skills required - Strong Python or Node engineer with production LLM pipeline experience: prompting, structured outputs, and evaluation loops - Google Workspace APIs (Gmail, Calendar, Drive) and HubSpot API, including OAuth and webhook plumbing - Entity resolution across messy identity data - Document generation (markdown → polished PDF) - Deployment and ops on Railway or equivalent — we hold full ownership of the code, instance, data, and keys, with a clean handover - Security-minded: this system touches sensitive deal and investor material; data handling, access, and deletion terms go in writing before anything connects to real mailboxes ## How we work - Direct, fast, async-first. Short written updates beat long meetings. - Feedback comes as redlines on real output — expect tight iteration, especially early - We supply everything you need: example briefings with annotations, working memo prompts, transcript exports, a real week of calendar, our HubSpot schema, and email access (starting from a forwarded subset to a temp inbox if you prefer to prove the sorting first) - Milestone structure: spec agreement → working prototype → backtest pass → shadow run → cutover - NDA required before we share the detailed spec and examples ## To apply, tell us 1. A similar system you've built (LLM + email/calendar/CRM), and your specific role in it 2. Your proposed stack and rough architecture in a few sentences 3. Availability and estimated timeline to a working prototype 4. Anything in the scope above you'd push back on — we'd rather hear it now 5. Any recommendations you would have that might work better in this automation, or any questions that would further help you