Skip to content

Sprint Plan

April 8 -- May 30, 2026. Engineering-first with a hard technical gate.

Delivered post-merge (2026-04-28)

Several deliverables originally pegged to the architecture-gate (Apr 29) and integration phases landed early. The current state:

  • Bridge + Coach engine wired: src/pitwall/__main__.py:/analyze returns rally-style pace notes from RuleCoach + LitertCoach, persists them to DuckDB. See api.md, ADR-010.
  • On-device LLM committed: Gemma 4 E2B via MediaPipe Genai (LiteRT-LM .task from litert-community/gemma-4-E2B-it-litert-lm). No cloud LLM tier. See ADR-012.
  • Frontend / backend boundary defined: backend owns LLM logic + system prompts; frontend renders only. See ADR-013.
  • Sonoma research integrated: 16 named track markers across 8 corners, per-corner Bentley + T-Rod tips, T-Rod voice in the Sonoma system prompt. 42 % of pace notes per lap reference real landmarks. See markers.md, sonoma_track_intelligence.md, trod_sonoma_session.md.
  • Session management API: /session/start, /sessions, /session/<id>, /session/<id>/end + coaching_notes table for off-track replay.

Items still on the original timeline below: field-test prep, regression suite, dry run.

Timeline

gantt
    title Trustable AI Sprint
    dateFormat YYYY-MM-DD

    section Milestones
    Technical Kickoff              :milestone, m1, 2026-04-08, 0d
    Architecture Review (HARD GATE):milestone, crit, m2, 2026-04-29, 0d
    Field Test - Sonoma Raceway    :milestone, crit, m3, 2026-05-23, 0d
    Sprint Wrap                    :milestone, m4, 2026-05-30, 0d

    section Phase 1 - Foundation
    Sensor fusion engine           :a1, 2026-04-08, 10d
    Antigravity pipeline Tx/Rx     :a2, 2026-04-08, 14d
    Confidence-annotated frame     :a3, 2026-04-08, 7d
    Racelogic + USB-CAN adapters    :a4, 2026-04-10, 10d
    Message arbiter                :a5, 2026-04-14, 7d

    section Phase 2 - Intelligence
    Gemma 4 hot path on Pixel TPU  :b1, 2026-04-18, 10d
    Pedagogical vector DB          :b2, 2026-04-18, 7d
    LitertCoach warm path on-device :b3, 2026-04-21, 10d
    Gold Standard baseline (AJ lap):b4, 2026-04-21, 5d

    section Phase 3 - Integration
    End-to-end integration test    :c1, 2026-04-28, 7d
    Signal Light HUD               :c2, 2026-05-01, 7d
    Audio UX tuning                :c3, 2026-05-01, 10d
    Per-pod persona tuning         :c4, 2026-05-05, 10d

    section Phase 4 - Field Prep
    Sonoma track definition        :d1, 2026-05-12, 5d
    Regression test suite          :d2, 2026-05-12, 7d
    Dry run (local track)          :d3, 2026-05-15, 5d
    Travel + setup at Sonoma       :d4, 2026-05-20, 3d
    Field test day                 :d5, 2026-05-23, 1d
    Debrief + documentation        :d6, 2026-05-24, 7d

The Hard Gate: April 29

No code, no track. Architecture review must demonstrate:

  • [ ] Sensor fusion engine running with Racelogic + USB-CAN data
  • [ ] Confidence-annotated frames flowing through the pipeline
  • [ ] Local DuckDB persistence working end-to-end
  • [ ] Gemma 4 inference on Pixel 10 TPU at <50ms
  • [ ] Message arbiter preventing conflicting coaching
  • [ ] At least 3 pedagogical vectors firing correctly on recorded telemetry
  • [ ] End-to-end: sensor → fusion → coaching → audio on a replayed session

Team Structure: F1 Garage Matrix

Vertical Pods

Each pod owns a complete user experience for one driver skill level.

Role Team 1: Beginner (Rental) Team 2: Intermediate (M3) Team 3: Advanced (Race Car)
Tech Lead Jigyasa Grover Hemanth HM Vikram Tiwari
Edge / Telemetry Madona Wambua Simon Margolis Austin Bennett (Mentor)
AGY Pipeline Mike Wolfson Taha Bouhsine Henry A Ruiz Guzman
Data Reasoning Adrian Catalan Vijay Vivekanand (Founder) Vinicius F. Caridá
UX / Frontend Rabimba Karanjai (Mentor) Aileen Villanueva Francisco Mere (Founder)

Horizontal Guilds

Cross-pod collaboration to build shared infrastructure once.

graph TB
    subgraph Edge Guild
        E1[Madona] --> SHARED_EDGE[Shared: Racelogic + USB-CAN<br/>adapters, sensor fusion,<br/>confidence annotation]
        E2[Simon]  --> SHARED_EDGE
        E3[Austin] --> SHARED_EDGE
    end

    subgraph Pipeline Guild
        P1[Mike]  --> SHARED_PIPE[Shared: DuckDB persistence,<br/>telemetry ingest,<br/>session management, bridge API]
        P2[Taha]  --> SHARED_PIPE
        P3[Henry] --> SHARED_PIPE
    end

    subgraph Reasoning Guild
        R1[Adrian]   --> SHARED_REASON[Shared: Pedagogical vectors,<br/>Gold Standard comparison,<br/>driver profile schema]
        R2[Vijay]    --> SHARED_REASON
        R3[Vinicius] --> SHARED_REASON
    end

    SHARED_EDGE --> PLATFORM[Shared Platform]
    SHARED_PIPE --> PLATFORM
    SHARED_REASON --> PLATFORM

    PLATFORM --> POD1[Pod 1: Beginner tuning]
    PLATFORM --> POD2[Pod 2: Intermediate tuning]
    PLATFORM --> POD3[Pod 3: Pro tuning]

Build horizontally first, tune vertically second.

Guilds build the shared platform (sensor fusion, Antigravity pipeline, pedagogical vectors) in Phase 1-2. Once the platform works end-to-end, each pod returns to vertical tuning: adjusting coaching language, thresholds, and persona for their specific driver level.

Taha's Role: AGY Pipeline, Team 2

Your scope on the Pipeline Guild:

Phase 1 (Apr 8-21): Build the Antigravity Pipeline

  1. Telemetry Persistence on Pixel 10: DuckDB-backed frame persistence with session management
  2. Bridge API on localhost: Flask endpoints for coaching, analysis, and session lifecycle
  3. Local reliability: DuckDB survives process restarts, no lost frames
  4. Burst format: Confidence-annotated frames in JSON, session metadata, driver level

Phase 2 (Apr 18-28): Connect to Reasoning

  1. Feed LitertCoach: Pass telemetry + Gold Standard + pedagogical vectors to on-device Gemma 4 E2B
  2. Return warm path coaching: Route LitertCoach response to arbiter in-process

Phase 3 (Apr 28 - May 15): Team 2 Vertical Tuning

  1. Tune for M3 / Intermediate: Adjust coaching cadence, LitertCoach prompt, coaching language for Team 2's BMW M3 and intermediate driver
  2. CAN configuration: Ensure USB-CAN adapter reads M3-specific CAN signals via DBC correctly

Phase 4 (May 15-23): Field Prep

  1. Sonoma dry run: End-to-end test on a local track or parking lot
  2. Regression tests: Verify pipeline on recorded Sonoma reference data

Deliverables

Completed (Pre-Sprint Data Analysis & Model Training)

  • [x] VBO parser — parses all 183 Racelogic .vbo files (535K frames, 14.9 hours)
  • [x] Track builder — auto-generates track definitions from GPS curvature. 3 tracks built: Sonoma (11 corners), Track 2 (9 corners), Track 8 (11 corners)
  • [x] Data analysis — 52 hot lap sessions profiled. Driving phase distribution: 43.7% cornering, 8.8% braking, 6.3% coasting, 2.2% trail braking
  • [x] Signal audit — 11 usable coaching signals, 7 broken/unmapped CAN signals documented
  • [x] LSTM v3 sequence predictor — trained on 140K sequences, tested on unseen track. Speed MAE: 3.3 km/h at 1s. Brake MAE: 2.7 bar at 1s.
  • [x] Phase classifier — XGBoost, 100% accuracy (labels are deterministic from features)
  • [x] Brake point predictor — linear regression, 15.9m MAE, each m/s adds 2.36m to brake zone
  • [x] Style fingerprint — K-Means 4 archetypes (aggressive, smooth, heavy braker, cautious)
  • [x] Sonic model v1 — hand-tuned audio cues (grip, brake approach, trail brake, throttle, coast)
  • [x] Sonic model v2 — LSTM-driven delta cues. Tested on Sonoma: fires speed_delta, brake_delta, lookahead, grip, corner score
  • [x] Simulator — replays VBO with real track data + LSTM model, exports labeled CSV
  • [x] Data documentation — 6 docs covering VBO format, signal reference, derived features, dataset overview, data quality
  • [x] Architecture docs — 12 pages + 9 ADRs for sprint edition
  • [x] Kaggle training pipeline — preprocessed data uploaded, training script tested on 2x T4 GPU

By April 29 (Architecture Gate)

  • [ ] Antigravity Tx/Rx working end-to-end
  • [x] Confidence-annotated frames flowing through pipeline (simulator demonstrates this)
  • [ ] Store-and-forward tested (simulate 5G dropout)
  • [ ] On-device warm path generating coaching via LitertCoach
  • [ ] Warm path response delivered to arbiter → earbuds
  • [x] Track definitions auto-generated for Sonoma (validated against real data)
  • [x] LSTM model predicting speed/brake/throttle 2 seconds ahead on unseen track

By May 23 (Field Test)

  • [ ] Full system running on Pixel 10 in an M3
  • [ ] Audio coaching audible and coherent via Pixel Earbuds
  • [ ] Signal Light HUD showing grip bars
  • [ ] Lap times computed from GPS crossing
  • [x] Corner report card framework built (corner scorer in train_models.py)
  • [ ] Driver profile updated from session data
  • [ ] System runs fully on-device without cloud dependency
  • [x] Pedagogical vectors defined with real Sonoma corner data

By May 30 (Sprint Wrap)

  • [ ] Session recordings from Sonoma field test
  • [ ] Post-session analysis comparing 3 pods
  • [x] Architecture documentation updated with field test findings (data analysis complete)
  • [ ] Reference architecture ready for Google I/O narrative

ADK Backend Upgrade — Shipped 2026-05-01

All phases delivered ahead of schedule. 358/358 tests passing. See ADR-019, ADR-020, ADR-021, and adk-agent-architecture.md.

gantt
    title ADK Backend Upgrade — SHIPPED
    dateFormat YYYY-MM-DD

    section Phase 1 - Conversation persistence
    conversations table + brief/debrief write     :done, e1, 2026-05-01, 3d
    GET /conversations endpoints                  :done, e2, 2026-05-01, 3d

    section Phase 2 - ADK agents (18 agents)
    15 tools + PitwallOrchestrator                :done, f1, 2026-05-01, 3d
    DebriefPipeline + BriefPipeline               :done, f2, 2026-05-01, 3d
    Runner + persistent sessions + KV cache       :done, f3, 2026-05-01, 3d
    DuckDB agent_traces + PitwallTracingPlugin    :done, f4, 2026-05-01, 3d

    section Phase 3 - Driver Q&A
    POST /coach/ask + POST /coach/ask/end         :done, g1, 2026-05-01, 3d
    GET /coach/agents (Vue PWA discovery)         :done, g2, 2026-05-01, 1d

    section Remaining backlog
    SSE streaming /coach/ask/stream               :h1, 2026-06-01, 7d
    Vue PWA paddock Q&A screen                    :h2, 2026-06-08, 14d
    ADK state scopes + plugin hooks               :h3, 2026-06-15, 7d

Phase 1 — delivered ✅

  • [x] conversations table in get_db() schema
  • [x] brief() + debrief() write narratives to conversations
  • [x] GET /conversations/<session_id>
  • [x] GET /conversations/driver/<driver_id>

Phase 2 — delivered ✅

  • [x] google-adk installed, src/pitwall/adk_tools.py (15 tools), src/pitwall/features/coaching/adk_agents.py (18 agents)
  • [x] PitwallOrchestrator(BaseAgent) with _classify_intent() deterministic routing
  • [x] DebriefPipeline: SequentialAgent → ParallelAgent([Highlights, Telemetry, Pedagogy]) → NarrativeAgentDebrief
  • [x] BriefPipeline: SequentialAgent → PedagogyAgent → NarrativeAgentBrief
  • [x] Runner + InMemorySessionService + run_adk() (canonical sync entry point)
  • [x] Persistent sessions per driver — KV cache reuse via session cloning, reset_driver_session() called at POST /session/start
  • [x] PitwallTracingPlugin(BasePlugin) + agent_traces DuckDB table — full agent + tool telemetry
  • [x] AgentMetaAgent — queries agent_traces for system observability
  • [x] All 7 ADR-020 audit fixes + 6 ADR-021 audit fixes shipped
  • [x] 358/358 tests passing

Phase 3 — delivered ✅

  • [x] POST /coach/ask — multi-turn Q&A, 6-turn rolling context, _qa_histories with 1-hour TTL
  • [x] POST /coach/ask/end — flush turns to conversations table
  • [x] GET /coach/agents — AGENT_REGISTRY for Vue PWA discovery

Remaining backlog

  • [ ] POST /coach/ask/stream — SSE streaming (RunConfig + StreamingMode.SSE)
  • [ ] Vue PWA paddock Q&A screen
  • [ ] ADK state scopes (user:, app:, temp:) — driver prefs persist across sessions
  • [ ] before_tool_callback per-agent SQL validation
  • [ ] LoopAgent for Q&A refinement (max 3 iterations)
  • [ ] Artifact storage for session PDF export