Roadmap

Full project timeline from the first legacy commit through the planned future. Shipped versions are solid entries; planned milestones carry a "Planned" badge; the side-repo ecosystem is treated separately at the bottom.

Timeline

Legacy phase

• Version: v0.1 to v0.5 Date: May 2025 Status: Shipped

  Legacy integration and project setup

  Third-year coursework iteration. Integrated the F1_Telemetry_Manager submodule, established the modular repository layout (src/, notebooks/, data/, legacy/), and verified seven FastAPI endpoint categories. A legacy_version branch preserves this work; active TFG development starts at v0.6.

Data and models

• Version: v0.6.0 Date: 2026-02-12 Status: Shipped

  Data engineering pipeline

  End-to-end pipeline from raw FastF1 telemetry to a clean, feature-rich dataset. Circuit clustering (K-Means k=4, 25 circuits), 48-column feature set over ~45,000 clean racing laps, and 2025 saved as a held-out test set. Dataset published to Hugging Face Hub.

  4 clusters ~45k laps 2023-2025
• Version: v0.7.0 Date: 2026-03-05 Status: Shipped

  ML foundation: lap time and tire degradation

  First two production ML models. XGBoost delta-lap-time predictor (N06) and a Temporal Convolutional Network for tire degradation with MC Dropout uncertainty (N07-N10, per-compound fine-tuning on SOFT / MEDIUM / HARD).

  Pace MAE 0.4104 s Tire MAE 0.708 s MC Dropout N=50
• Version: v0.8.0 + v0.8.1 Date: 2026-03-13 Status: Shipped

  Extended ML models: overtake, safety car, pit stop, undercut

  Four additional predictors. LightGBM overtake classifier on 28,494 labeled pairs (N11-N12), soft safety-car prior (N13-N14), HistGBT quantile regression for pit duration (N15), and LightGBM undercut success scorer (N16). The Causal TCN alternative (N12B) archived as a documented negative result.

  Overtake AUC-PR 0.5491 SC lift 1.67x Pit MAE 0.487 s Undercut AUC-ROC 0.7708
• Version: v0.8.2 Date: 2026-03-22 Status: Shipped

  NLP radio processing pipeline

  Full team-radio NLP stack. Whisper ASR (N18), RoBERTa sentiment (N20, 0.84 accuracy), SetFit intent classification (N21, 5 classes), BERT-large NER for F1 entities (N22), deterministic RCM parser (N23). Unified inference entry point in N24 at GPU P95 latency of 59.4 ms.

  Sentiment 0.84 P95 59.4 ms GPU

Agents and distribution

• Version: v0.9.0 Date: 2026-03-17 Status: Shipped

  src/ extraction, CLI simulation, radio corpus

  Seven agent entry points extracted to importable src/agents/ modules. Headless CLI simulation (f1-sim) with Rich Live rendering and no-LLM mode. OpenF1 radio corpus pipeline (529 MP3s, 48 parquets) with Whisper JSON cache. Lazy Hugging Face data download on first run.

• Version: v0.10.0 Date: 2026-03-22 Status: Shipped

  Multi-agent system: N25 to N31

  Six LangGraph ReAct sub-agents coordinated by the N31 Strategy Orchestrator's three-layer pipeline: MoE-style dynamic routing, Monte Carlo simulation over 500 samples ranking four strategy candidates by risk-adjusted expected outcome, and LLM synthesis producing a 14-field StrategyRecommendation. Bahrain 2025 end-to-end demo.

  6 agents + 1 orchestrator MC N=500 MoE routing
• Version: v0.11.0 Date: 2026-03-30 Status: Shipped

  RAG system over FIA regulations

  Retrieval-augmented generation grounding strategic decisions in FIA Sporting Regulations 2023-2025. BGE-M3 embeddings (1024-dim), Qdrant local vector store, 2,279 indexed chunks. Exposed as the query_rag_tool LangChain tool imported by the orchestrator.

  2,279 chunks BGE-M3 1024d Scores 0.62-0.76
• Version: v0.1.1 Date: 2026-04-09 Status: Shipped

  R1: CLI wheel release

  First distribution artifact. Wheel f1_strat_manager-0.1.1-py3-none-any.whl on GitHub Releases. Both f1-strat and f1-sim entry points verified. Installable via uv tool install git+<repo>.

• Version: v0.12.0 Date: 2026-04-15 Status: Shipped

  Interfaces and distribution: FastAPI, Streamlit, Arcade

  Wired the multi-agent system into the FastAPI backend and FastMCP. Streamlit pages for race analysis and live strategy cards with inline Plotly charts. Arcade MVP: three windows from one command (pyglet 2D replay, PySide6 strategy dashboard, live telemetry grid). Voice chat rewritten with Whisper ASR and edge-tts.

  3 surfaces 7 MCP tools TCP 10 Hz stream

Stable release

• Version: v1.0.0 Date: 2026-04-20 Status: Shipped

  Final release: code freeze and three-surface distribution

  First stable release. Consolidates the Arcade MVP, the full seven-model ML stack, and N25-N31 with FIA RAG into a single tagged production release. Three install paths: CLI wheel, f1-arcade via uv tool install, and Streamlit + FastAPI Docker Compose.

• Version: v1.1.0 Date: 2026-05-11 Status: Shipped

  Benchmark suite and thesis evaluation

  Four standalone benchmark scripts covering pace, Whisper latency, six sub-agent latency, and the NLP pipeline on CPU and GPU. N33 precision-recall sweeps for overtake, safety car, and undercut models. The shared f1-eval rag command evaluates the production retriever over 30 ground-truth queries. Full English localization of strategy notebooks.

• Version: v1.2.0 to v1.3.1 Date: 2026-05-12 Status: Shipped

  Documentation site launch and custom domain

  React docs site (plain React.createElement scripts, no JSX, no Babel, no build step) published under docs.f1stratlab.com. Full architecture pages, agent API reference, arcade quick-start, CI/CD narrative, and graph-based page discovery. Five drawio architecture diagrams. Brand theme aligned with the F1 StratLab purple palette.

• Version: v2.0.0 Date: 2026-07-21 Status: Shipped

  Modern frontend: the post-race UI moves to a React web app

  The Streamlit post-race interface is fully replaced by a React web app (React 19, Vite, TypeScript, Tailwind, ECharts, TanStack Router/Query, Zustand) across six surfaces: dashboard, comparison (60fps replay), model lab, strategy, race analysis, and a streaming chat that renders each tool result inline. The FastAPI backend, the six agents, the orchestrator and the models are unchanged: a presentation-layer swap, not a rebuild.

• Version: v2.1.0 to v2.5.1 Date: 2026-07-28 Status: Shipped

  Decision memory, and a round of correctness hardening

  The orchestrator gained a per-race decision memory: a caller-owned accumulator rendered into the Layer 3 prompt, wired through the CLI, the arcade and the backend, so a recommendation can say what changed since the last lap and why. Alongside it, a run of correctness work on the Monte Carlo and the agents, including banding threat_level on the served calibrated scale rather than on raw-scale operating points, pricing the lap-number neutralisation union in all three places it lives, and breaking an exact scenario tie by a stated rule instead of by dictionary insertion order. Full detail in the changelog.

• Version: v2.6.0 Date: 2026-09-03 Status: Shipped

  PITWALL, a designed Arcade, and a startup that stops making you wait

  The trackside surface is now PITWALL: two desktop windows built in React and hosted in a platform webview, a DATA window with the timing tower, the bests panel, the race-pace grid, the track ring and the radio feed, and an AGENTS window that shows the decision and the six consoles behind it. They run alongside the pyglet replay rather than replacing it, and the lap_state contract and the agents are unchanged. The replay itself got a design pass over the launch menu and the circuit layout, and every one of the 70 rounds of 2023 to 2025 now loads from the menu where only one did.

  The AGENTS window also gained a PIT EXIT card, which answers the one question a pit wall asks before every stop: if we box on this lap, what position do we rejoin in, and which cars land either side. It reads P1 → P3 under an if we box now header, with the car ahead and the car behind named and the gap to each. The number is the one the projection layer is graded on, 86.1% within one position over 552 real green-flag stops of 2025, because it is computed at the same two-lap horizon the ground truth measures rather than at the five-lap one the strategy scoring uses. RADIO gave up one of its two columns to make room, so the bottom row now runs three columns wide.

  Startup and per-lap cost came down with it. f1-sim --help went from 12.6 s to 3.6 s and a five-lap offline run from 15.6 s to 5.5 s, because importing an agent no longer drags the LLM stack in behind it; N25's confidence interval went from 660 ms to 5.9 ms a lap by scoring one perturbed block instead of two hundred single rows, returning the same numbers; loading an already-cached race got faster with the garbage collector held off during the read, and the pit-wall windows no longer pay 300 ms for the race on the first request they block on. Full detail in the changelog.

• Version: v2.7.0 Date: 2026-09-29 Status: Shipped

  Provider handling, RAG refresh and replay fixes

  No-LLM PLAN now distinguishes an absent future stop from an unknown plan. Requests that omit provider keep the configured default. Season-scoped, article-aware FIA retrieval covers the 2026 rules, and replay caches load faster with lower memory use. The deterministic scorer and both five-lap horizons are unchanged after the #715/#724 evidence review. The release notes list the complete changes.

Planned milestones

• Version: v2.8.0 Status: Planned

  Rival Agent: anticipate each rival's next move

  A new, additive LangGraph node that predicts each nearby rival's next strategic move (pit window, compound, undercut / overcut) and feeds it to the orchestrator. Recommendations move from reactive to anticipatory. The six existing agents are untouched. Validated by ablation against real 2024-2025 pit-stop outcomes.

• Version: v3.0.0 Status: Planned

  Live race inference and 2026 regulation adaptation

  Real-time OpenF1 WebSocket ingestion. The lap_state contract is unchanged, so agents and the orchestrator do not change. Adaptation to the 2026 technical and sporting regulation: re-cluster circuits, re-label compounds, add drift monitoring.

Post-TFG project proposals

The post-TFG plan proposes four projects outside the core release train. Their repository boundaries and release cadence are not committed.