1 / 21
←→ navigate
01

Meet Lovelace

Your AI Field Operations Assistant

39+
Schools
30+
AI Tools
12
Team Members
🇰🇪 Complementary Schools Foundation Kenya

Presented by Mulili · June 15, 2026

02
Architecture

What Is Lovelace?

Agent + Mixture of Specialized Models

🏆 TEACHER MODELS (Distillation)
DeepSeek-V4 · GLM-5.2 · MiniMax-M3 — frontier models used for training data generation
⬇ distilled into
⚡ MODEL MIXTURE — Shared H200 GPU Pod
Qwen3-8B & Gemma-4-12B (LLM) · Kimi-K2.7 (Code) · WhisperX (ASR) · Kokoro (TTS) · NeMo (Diarization)
⬇ powers
🧠 LOVELACE AGENT
30+ Tools · RAG Memory · Orchestration
Fully Multimodal
📝 Text 📸 Pictures 🎙️ Voice 🎬 Video 🗣️ Speech
03
Models

Model Architecture Explained

A Mixture of Best-in-Class Models

  • Frontier models proved too expensive & unreliable for the token budget required to make Lovelace a tool that does useful work for CSF
  • Replaced single-model approach with a mixture of specialized models — each best-in-class for its task
  • All models run simultaneously on a single NVIDIA H200 (141GB) using our own internal servers
🧠 Qwen3-8BCore LLM for conversation, RAG queries, field note analysis, and tool orchestration
💎 Gemma-4-12BLarger LLM for complex reasoning — meeting notes generation, report writing, deep analysis
💻 Kimi-K2.7 CodeSpecialized code model for the lovelace-code service — generates, reviews, and debugs code
🎙️ WhisperXSpeech recognition with word-level timestamps — transcribes meetings and voice notes
👥 NeMo MSDDSpeaker diarization — identifies who said what in multi-speaker recordings
🗣️ Kokoro NEWText-to-speech — gives Lovelace a voice to speak and summarize meeting notes aloud. Launching today!
All six models share a single GPU pod — zero idle cost when not in use, instant spin-up when needed via RunPod.
04
Platform

The Agent — 7 Modules

Go 1.24 · React 19 · PostgreSQL + pgvector · Redis · Nginx · systemd

🤖 lovelace-botTelegram bot in Go — RAG engine, 30+ tools, 15+ intents, 6 background workers, dual-LLM failover
🎤 lovelace-meetGoogle Meet note-taker with speaker diarization, plus new TTS summary via Kokoro
💻 lovelace-codeCoding module powered by Kimi-K2.7 — being introduced to Caiden next Wednesday to drive csfkenya.com development
📊 lovelace-dashboardUI + API at dashboard.csfkenya.com — schools, KPIs, OCR, media, calendar, user management
📱 lovelace-miniappsTelegram WebApp views — field notes, action items, school profiles, media gallery, spending, calendar (10 mini-apps)
⚙️ lovelace-inferenceGPU inference orchestrator — manages the Mixture of Models (MoM) on RunPod H200 pods for ASR, LLM, TTS, and OCR
📈 lovelace-data-analysis NEWIntegrates with School Management Systems (ShuleOne & Smart School) recommended by CSF — ingests enrollment, attendance, and fee data directly. Insights are already promising, but integrating this module is what has delayed Lovelace coming online the last 2 weeks.
47
DB Tables
90+
Specs Written
30+
AI Tools
6
Workers
05
Why

Why All This Complexity?

One word: Tokens

For Lovelace to be a useful tool — both short and long term — CSF needs sustainable token throughput: the ability to process large volumes of AI work reliably, every day, without runaway costs.
🧩 TokenThe smallest unit an AI model reads and writes. Every word you send, every answer you get, every document analysed — costs tokens. Think of tokens as the "fuel" that powers every AI interaction.
📏 ContextThe amount of information an AI can "see" at once. A bigger context window means Lovelace can read an entire meeting transcript, all 39 schools' data, and your question — all in one go. Without enough context, the AI gives shallow answers.
📚 RAG (Retrieval-Augmented Generation)Instead of stuffing everything into context, Lovelace searches its database for the most relevant information first, then feeds only that to the AI. Accurate, grounded answers — without burning tokens. I learnt this well from reverse engineering the API calls of the Antigravity tool that Alister recommended.
🔄 Agentic LoopLovelace doesn’t just answer questions — it takes actions. It can look up a school, check KPIs, schedule a meeting, and send a summary — all in one conversation. Each "loop" uses tokens, which is why efficient throughput matters so much.
The Bottom Line

Frontier models like Claude and GPT charge per token. At CSF’s scale — 39 schools, daily field notes, weekly meetings, thousands of documents — that bill grows fast. Our own models give us unlimited tokens at a fixed cost.

06
Journey

How Lovelace Has Evolved

From Short Memory to Long Memory

v0 · Sep–Oct 2025
Basic RAG chatbot — answered questions about schools but forgot everything between sessions.
v1 · Nov–Dec 2025
Session persistence, school context, field notes ingestion. Started distilling base model.
v2 · Jan–Feb 2026
Meeting notes pipeline, action items, SMS/school data integration, spelling corrections.
v3 · Mar–Apr 2026
Group visit awareness, notification intelligence, story arc tracking, OCR pipeline.
v4 · May 2026
Voice transcription, broadcast hooks, Dashboard v1, WhatsApp integration ready.
v5 · Jun 2026
Mixture of Models (MoM) architecture, lovelace-code module, SMS data-analysis (ShuleOne & Smart School), Kokoro TTS for meeting summaries, Zoom meeting support.
Currently at ~35% of the full vision — accelerating. Every sprint delivers more than the last.
07
Features

Telegram Bot

What Lovelace Can Do Today

🎙️ Voice Field NotesRecord observations hands-free, auto-transcribed & structured
📊 KPI IndicatorsEnrollment, fee collection, teacher retention — queryable per school
✅ Action ItemsExtract & track from meetings and emails, weekly carry-forward
📸 Photo/Video UploadsAuto-linked to schools via GPS + captions, OCR-capable
📍 Location AwarenessGPS-based school detection, group visit broadcasting
💬 WhatsApp IntegrationFully built, pending Meta Business registration
📧 Email IntegrationCC'd email → action items, automated reminders
💰 Financial TrackingSchool spending queries, cross-school comparisons
📈 Impact ReportsAuto-generated school impact summaries
📝 Field Note DigestsDaily 3AM digest, auto-enriched and embedded in RAG
🗺️ School Directory40+ schools queryable by name, location, contacts
🔄 Self-HealingAuto-deploys, rebuilds binary, migrates DB — zero downtime
08
Intelligence

Smart Notifications

Proactive Field Support

  • 🔔 Pre-visit briefings with RAG-sourced context — action items, enrollment data, pending follow-ups
  • 📡 Group visit broadcasting — one person sets context, entire team syncs automatically
  • 🧠 Context-aware — knows which school you're at, what you've documented, what's pending
  • 🎯 Personalized per team member — each person gets their name, their action items
  • 🕐 Scheduled delivery — notifications fire at the right time
  • 📧 Email OCR worker — scans incoming emails for documents to OCR automatically
  • 🌙 Midnight sweep — daily cleanup of sessions, stale data, orphaned jobs
Notifications transform Lovelace from a reactive tool into a proactive assistant that anticipates your needs.
09
Pipeline

Meeting Notes

Friday Meeting Pipeline — 10 Phases

  1. Join & Record — Connects via Google Meet and records the full audio. Lovelace captures everything spoken in the meeting
  2. Context Loading — Loads spelling corrections, team roles, 40+ school names, story arcs
  3. Raw Extraction — Reads transcript in 150-line chunks. Captures everything
  4. Report Generation — Warm, quote-heavy, detail-obsessed report
  5. Audit — Re-reads entire transcript to catch anything missed
  6. HTML Generation — Styled with callout boxes (🟢 wins, 🟠 issues, 🔴 urgent)
  7. Cross-Check — Verifies plain text ↔ HTML parity
  8. User Review — Presents both versions, flags new names
  9. Corrections — Applies feedback to both files
  10. Commit — Pushes to git with updated pointers
Delivered — Rolling Out Tomorrow

✅ 🎙️ Meeting notes in ALL calls — not just Fridays. Lovelace joins every call, takes notes, and extracts action items automatically.

✅ 🔒 Confidentiality guardrails — "don't tell X" = excluded from report.

10
Dashboard

The Dashboard

Replacing Notion

The old Notion system couldn't handle the volume and complexity of CSF data. The dashboard is v1 — it only gets better.
  • 📊 Dashboard Home — Stats cards, charts, calendar widget, activity feed
  • 🏫 Schools — List view + interactive map, detail pages with tabbed interface:
        Overview · Field Notes · Media (HLS streaming) · KPIs · Tasks
  • 📄 OCR Pipeline — Upload page, job queue, detail view with extracted data
  • 📊 KPI Dashboard — Edit/delete modals, cross-school comparisons
  • 📅 Calendar — Full calendar with month grid, timeline sidebar, event forms
  • 👥 User Management — Team profiles, role-based access
  • 🔐 Auth — Telegram-based SSO, deploy overlay for maintenance
11
Data

OCR & KPI

From Paper to Data

OCR Pipeline — Key Capability
  • Take a photo of any school document → Lovelace extracts the data
  • Enrollment forms, fee receipts, budgets, KCPE results — all understood
  • Data flows into KPI indicators automatically — no manual entry
  • Invoices & receipts tracked to keep tabs on spending vs KPI data
  • Goal: schools go fully paperless via Lovelace

📊 KPI Dashboard

📈
Enrollment
💰
Fee Collection
👩‍🏫
Teachers
🏗️
Infrastructure
Engaging Victor & Ken for SMS API access — real-time KPI data from all schools starting next week.
📱 DEMO: Let’s do a live OCR right now — take a photo of any receipt or document and watch Lovelace extract the data in real time.
12
Delivered

What's New Since Last Time

All features rolling out tomorrow at 10am BST

  • ✅ 🎙️ Meeting notes in ALL calls — Lovelace joins every call, takes notes, extracts action items automatically
  • ✅ 📝 Dashboard: View meeting notes — Read and search all transcripts from the web
  • ✅ ✏️ Dashboard: Edit field notes — Correct and annotate from the web
  • ✅ 📊 Dashboard: Enhanced KPI views — Trend charts, school comparisons after SMS integration
  • ✅ 💬 WhatsApp launch — Pending Meta Business registration (internal use only for now)
  • ✅ 🎬 Presentations on the fly — Generate polished slide decks instantly, like this one
New Feature

💻 lovelace-code — A dedicated coding module powered by Kimi-K2.7. Being introduced to Caiden next Wednesday to drive the development of csfkenya.com.

Every item from the last presentation has been delivered. The system compounds — every field visit and meeting makes Lovelace smarter.
13
Medium Term

Next 4 Months

Deep Intelligence & Voice

Voice Conversations

🗣️ Lovelace speaks and listens — in meetings, on calls, and 1-on-1 with staff. Not a novelty voice like Siri — a serious, respectful conversational tone that understands context, remembers history, and speaks with cultural awareness.

  • 📄 Supplier & Pricing Intelligence — After OCR of all CSF receipts and invoices (2024–present), Lovelace extracts every supplier, tracks pricing trends across similar infrastructure projects (solar, water tanks, construction), and flags anomalies
  • 📊 Cross-school cost benchmarking — "School X paid 40% more for the same solar installation" — surfaced automatically
  • 🧾 Tax Filing (Kenya & UK) — Ability to file tax returns in both jurisdictions, pending CSF admin approval. Extensible to other schools in the long run — taxes are taxes. Trial runs have already started
  • 🔗 Financial Management Integration — Tight integration with the financial management system Victor-Brilliante is developing
14
Roadmap

Term 3 2026

Pilot: Lovelace for Schools

  • Introduce Lovelace to 2–3 pilot schools (at CSF's discretion)
  • Start as a "toy" — directors can:
    • Check their KPIs
    • See how their school performs vs peers
    • View their own field notes and action items
  • Low-risk exploration — no operational dependency yet
  • Gather feedback on what school directors actually want
The pilot phase lets us learn what directors value most before we scale.
15
Roadmap

Term 1 2027

Monitoring: Day-to-Day Operations

Expand Lovelace to actively monitor school operations:

💰 Fee CollectionDaily tracking and trend analysis
👩‍🏫 Teacher AttendanceAutomated attendance awareness
📊 Enrollment AlertsReal-time change notifications
🏗️ InfrastructureIssue flagging and tracking
  • System provides proactive insights to directors and CSF simultaneously
  • Still supervised — CSF team reviews and validates
16
Roadmap

Term 2 2027

Full Deployment: All Schools

  • Deploy across all CSF schools
  • Daily support and coordination without straining the team
  • Automated daily check-ins, issue escalation, resource coordination
⚠️ Election year context: Kenya elections create chaos — school disruptions, staff turnover, community tensions. Lovelace handles the coordination load so the team can focus on what matters.
17
Vision

Election Year 2027

Lovelace as Crisis Coordinator

During Kenya's election period, things get unpredictable:

  • School closures, staff displacement, community tensions
  • Communication becomes harder, travel becomes risky

Lovelace Should Be Able To:

  • 🌍 Coordinate across all schools with minimal hallucinations
  • 📡 Real-time situational awareness without physical presence
  • 🛡️ Reduce strain on the team during the most chaotic period
By this point: battle-tested model, 18+ months of field data, deep school context across all 39+ schools.
18
Infrastructure

The DGX Spark RENTED ✅

From Cloud to On-Premise · nvidia.com/dgx/spark

128GB
Unified Memory
1 PFLOP
AI Performance
$4,699
MSRP
1.2 kg
Weight

Grace Blackwell GB10 Superchip · 20-core ARM CPU · 4TB NVMe SSD · Wi-Fi 7 · 100GbE · fits on a desk

Why DGX Spark vs Frontier Models
  • 🔒 Security & Privacy — All CSF data stays on-premise. No school records, student info, or financial data ever leaves the building
  • 📈 Grows with CSF — The custom model improves with every interaction. Two units link together for 405B parameter models as needs scale
  • 💰 Zero recurring AI cost — Frontier models charge per-token forever. DGX Spark is a one-time purchase — after that, just electricity
  • ⚡ Speed & Reliability — No internet dependency for inference. Works even when connectivity is poor in rural Kenya
  • 🎯 Full Precision — Running bf16 (Brain Float 16) weights — the native training format, not compressed or quantized. Full quality, zero degradation
~$38K
Dev Cost to Date
$48K
Original Budget
~$0/mo
After DGX Spark
✅ Already rented and ready to be shipped to Kenya with help from Felix. One-time hardware cost, then unlimited AI for CSF — forever.
19
Long Term

The Long Term Vision

An Ever-Present Companion 😊

Lovelace’s long-term goal is to be an ever-present companion for CSF staff, school directors, and teachers — helping schools on a day-to-day basis with the capabilities of a PhD-level accountant, mathematician, programmer, and storyteller.

Future Risks to Watch
  • ⚠️ Hallucinations — AI models can confidently state incorrect information. Mitigation: tight integration with CSF’s own verified data and more intensive training on school-specific context. The more real data Lovelace has, the less it needs to guess
  • 🧠 AI Psychosis — An emerging mental health trend where people become overly dependent on AI models, especially when the AI consistently “out-meets” expectations or inadvertently gaslights users. We will address this proactively if it ever arises — clear boundaries, transparency, and human-first design
Awareness of these risks now means we can build guardrails before they become problems.
20
Transparency

Code Sharing & Audit

Open Books, Open Weights

  • 🔒 The agentic code lives in a private GitHub repository — I’m willing to share it with anyone for audit at any given time
  • ⚙️ The agentic code is written in Go (Golang)
  • 🧠 The base model weights are hosted on our own internal server alongside the inference infrastructure (vLLM)
  • 📊 Model weights are full precision bf16 (Brain Float 16) — native training format, not quantized or compressed. This is the real thing, not a degraded copy
Open Invitation

I am fully open and transparent — happy to share the codebase and model weights with Alister’s team or PG’s team for an independent audit, so they can verify every claim about what Lovelace does.

✅ Nothing is hidden. The code, the model, the data pipeline — all available for inspection on request.
21

Thank You

"Lovelace is 25% of the vision — but the foundation is solid"

  • 🧠 The base model gets smarter with every interaction
  • ⚡ The agent gets more capable with every sprint
  • 📊 The data gets richer with every school visit

Questions? 🙋