PACE Compass — Pierre's Department & Career Dashboard

Department at a glance

Five overlapping workstreams. Not one thing.

PACE certification on track

101 ✓ · 201 ✓ · 301 Phase 13 prep

5/6 modules83%

Data-Workbench split

3 locations: skill, pace-local, standalone repo. Consolidation candidate.

Insight pipelines 1/4 warm

panthera-titration warm · email-analysis · genai-naep · session-analytics cold

Verified research spec'd

Perplexity pipeline designed, working, not yet wired to brain DB

Knowledge / brain live

27 learnings · 2 CF deploys · brain DB v2.sessions streaming

Overall milestone progress

0 / 0 done0%

Live deploys: pace-study.pages.dev · data-workbench.pages.dev · pace-compass.pages.dev (this one) — shared PIN via pace_pin_auth_v1.

Milestones — tick as you go

Persists per-device via localStorage. Each group shows its own %.

PACE certification

CAM_DS_101 Applied Statistics
Complete · course scraped to canvas-export/
CAM_DS_201 Supervised ML
Complete · pace-course2 portfolio + LinkedIn post drafted
CAM_DS_301 NLP — rubric 48/48
All rubric items covered in v3 pipeline
CAM_DS_301 walkthrough deployed
pace-study.pages.dev live
CAM_DS_301 submission notebook (basic)
Run patched notebook top-to-bottom on A100 · sync back to repo
CAM_DS_301 submitted
Deadline 2026-04-27 17:00 UK

8-project portfolio (DeepMind roadmap)

1. Education Data Pipeline
in_progress per yaml · but: data-workbench infra already ships this — rename?
2. Student Success Predictor
Needs Sparx task-attempt data access
3. Student Error Type Classifier
Multi-class; random-forest / XGBoost
4. Learning Trajectory Model
Time-series / LSTM — DeepMind-shaped
5. ML Evaluation Pipeline
Cross-validation, metrics for educational ML
6. Adaptive Learning Recommender
RL / recommender systems
7. Interpretable ML for Teachers
SHAP / LIME explanations
8. Education ML Research Paper
Combine PACE techniques with Sparx data

Vertex AI learning NEW

Detailed roadmap on the Vertex AI tab. Checked items sync.

GCP + Vertex AI account / free tier
Billing alerts set ≤ £5/month
Vertex AI Studio walkthrough
Prompt design + model comparison in console
Model Garden: deploy an endpoint
Pick one model · deploy · call from Python SDK
Gemini via Vertex (vs direct API)
Grounding, function calling, system instructions
Vector Search (formerly Matching Engine)
Embed · index · query — brain DB doc_chunks equivalent
Vertex AI Pipelines (KFP)
One real DAG — e.g. the PACE emotion pipeline
Custom training job
BYO container · TPU/GPU pricing sanity-check
Agent Builder: one agent with tool use
Compare to current brain/persona architecture
Portfolio project: port PACE → Vertex
Batch prediction job for emotion classifier · publishable

Department hygiene

Commit uncommitted work
pace-nlp-project + brain-vault · workbench tooling + patched notebook + learnings
Wire data-workbench guard hook
One line into acebuddy/scripts/file-guardrail-hook.sh
Consolidate workbench (pace-local → standalone)
After 2026-04-27 submission
Wire perplexity-api-test → brain DB
Tier-classified writes to learnings + research_outputs
Email-analysis pipeline — first pass
127k records in brain DB; attention pattern signal
Reconcile career.yaml vs sabbatical state
See Career tab
Scrape Career Toolkits + Owning Your Career courses
Canvas IDs 519 + 520 · already paid for

Career plan — current yaml

Target

DeepMind Research Engineer — AI for Education

Timeline: 2030 · Angle: PhD in learning mechanisms + Sparx data at scale

Career state — answered 2026-04-18: 0.8 FTE Sparx since Jan 2023 (both realities live)

career.yaml confirmed accurate: current_role: "School Success Coach at Sparx Learning (0.8 FTE)". Start date added (current_role_start: "2023-01") + last_reviewed: "2026-04-18". Not transitioning — both Sparx (present) and DeepMind 2030 (long-range) coexist.

On sabbatical from Sparx (yaml still right) Sparx chapter closed (yaml stale — update) Other / complicated (tell me in next session)

Saves locally. Next session I'll read this and update yaml accordingly.

DeepMind 2030 — answered 2026-04-18: still live

Still live — keep roadmap as-is Dormant — not ruled out, not current focus Shifted — newer target (tell me in next session)

Portfolio — from yaml

#	Project	Status	Career score
1	Education Data Pipeline	in_progress	—
2	Student Success Predictor	planned	logistic-regression: 9, nn: 10
3	Student Error Type Classifier	planned	random-forest: 8, xgboost: 9
4	Learning Trajectory Model	planned	neural-networks: 10
5	ML Evaluation Pipeline	planned	evaluation-metrics: 8
6	Adaptive Learning Recommender	planned	RL / recommender systems
7	Interpretable ML for Teachers	planned	explainability
8	Education ML Research Paper	planned	publication

Honest observation

Your actual current IP: cross-environment data tooling, safe-execution discipline for LLM workflows, verified-research pipelines, brain+persona architecture. The DeepMind-shaped yaml reads like a 2024 pre-LLM-era portfolio. Not wrong — but the frontier moved while you were building. Worth a fresh pass on whether "DeepMind 2030" or "AI-infra principal / staff at a well-funded research-adjacent team" is the better-aimed shot. Park or pick.

LinkedIn templates in yaml

Three post templates exist: technique_insight, sparx_application, portfolio_update. Hashtags include #DataScience #MachineLearning #EdTech #AIinEducation #CambridgeDataScience.

Vertex AI learning track

GCP's managed ML platform. Gap in current stack (AWS/local/HF). Cloud-neutral pedigree + Google-adjacent for DeepMind line. Budget-first approach.

Cost guardrail: set billing alert at £5/month before any experimentation. Vertex compute can burn fast (TPU/A100 per-minute). Default to Model Garden + Gemini API (token-priced) for most learning. Only spin custom training jobs when you know the run duration.

Roadmap

1. Foundations ~2h

Create GCP project, enable Vertex AI API
Free tier: $300 credit 90 days · set hard budget cap
Install + authenticate gcloud CLI
gcloud auth application-default login
Vertex AI Studio — prompt design console
Compare Gemini 2.5 Flash vs Pro on a PACE review classification task

2. Model Garden ~3h

Deploy one endpoint from Model Garden
e.g. Llama-3.1-8B or Gemma-7B · note cost/hr of the smallest machine
Call endpoint from Python SDK
google-cloud-aiplatform
Undeploy endpoint (don't forget — billing)
Endpoints accrue cost while running, even idle

3. Gemini via Vertex ~2h

Gemini 2.5 basic call (pay-per-token)
Cheaper than endpoint deploys for learning
Grounding with Google Search
Built-in retrieval for freshness
Function calling / tool use
Structured output via schema

4. Vector Search ~4h

Generate embeddings (text-embedding model)
Compare to OpenAI ada for cost + quality
Create Vector Search index
Brain DB doc_chunks equivalent · ~1k docs to start
Query index from Python
Nearest neighbour + metadata filter

5. Pipelines (Kubeflow) ~6h

Hello-world pipeline (2 components)
KFP SDK basics
Port PACE emotion pipeline as a DAG
load → clean → classify → report · portfolio-publishable

6. Custom training ~4h

BYO training container
Dockerfile + training script
Hyperparameter tuning job
Small search, single region, time-capped

7. Agent Builder ~3h

Build one agent with two tools
Compare UX + cost to your brain/persona system

8. Portfolio project capstone

PACE emotion classifier → Vertex batch prediction
Deploy · schedule · cost-report · write-up for LinkedIn
Full PACE v3 pipeline as a Vertex Pipeline
Rubric-to-cloud port · DeepMind-relevant narrative

Good resources (unopened)

cloud.google.com/vertex-ai/docs — official
Google Cloud Skills Boost (formerly Qwiklabs) — hands-on with sandbox projects
github.com/GoogleCloudPlatform/vertex-ai-samples — copyable notebooks
"Prompt design with Gemini" + "Vertex Vector Search" learning paths on Skills Boost

Open threads — set priority

Edit priority numbers → auto-sort. 1 = ship next. Saves locally. Colour: 1-3 red, 4-6 amber, 7+ green.

Miss a thread? Priorities are yours — I'll pick up whichever sits at #1 next session.

Panel — next block of time

Fork: 9 days to PACE · ambiguous career state · uncommitted work · new Vertex track.

Architect

Consolidate workbench into standalone repo now. One night's work. Commit everything, wire guard hook, lift to ~/projects/data-workbench/. Pays forward into every future insight pipeline — including Vertex.

Risk: not on the rubric, not before deadline.

Cognitive Load Specialist

Commit everything today (30 min — protects state). Defer consolidation until after 2026-04-27. Start Vertex AI only after submission. You have a known ADHD pattern of starting rebuilds mid-project. Submit PACE clean, then refactor, then reopen career.

Risk: uncommitted discipline lingers 9 more days.

Behavioural Product Designer

Radios answered 2026-04-18: career state = "other/complicated" (both realities live — 0.8 FTE Sparx since Jan 2023 and DeepMind 2030 as long-range aim), DeepMind 2030 = live (north star). career.yaml updated with start_date + last_reviewed. No transition underway; the complexity is holding both simultaneously.

Risk resolved: career.yaml was accurate, not stale.

My read (subtractive)

Order:

Finish PACE submission (9 days to 2026-04-27 17:00 — A100 run complete, notebook at basic/basic_notebook.ipynb; submit via Canvas assignment 3354 at fourthrev.instructure.com as .ipynb + report.md or zip)
Then Vertex AI track · start with Foundations (£0 budget, 2h) + Gemini-via-Vertex (token-priced, small)
Then wire perplexity → brain DB
Status check on cold pipelines: email-analysis / genai-naep / session-analytics — alive or parked?

Retire: urge to restructure the department before submitting. Career.yaml update is all the restructuring this month needs.

Retired 2026-04-18: workbench-lift merge, guard-hook wiring, pace-compass git init — all shipped.

Quick actions

What I'm tracking for you between sessions

Tick state, hold-list priorities, career radio answers all live in localStorage on this device. Use Export state to back up. In next session I'll read this before planning.