Embodied AI Architect — Feature Inventory & Implementation Roadmap

Complete · updated 2026-10-30 · embodied-ai-architect/docs/plan/feature-inventory-and-implementation-roadmap.md

Context

The Embodied AI Architect is an agentic AI platform for autonomous HW/SW co-design of embedded AI systems. After 3 months of development (Dec 2025 – Feb 2026), significant infrastructure exists: 6 core specialists, KPU RTL generation, multi-objective optimization, codebase analysis, 7 demos, and 560+ tests. However, critical gaps remain — the co-simulation validation pipeline was skipped, CLI commands have stubs, distributed backends aren’t wired, and the optimization loop (Phase 2) is framework-only.

Goal: Enumerate all features/subfeatures, organize into an 18-month roadmap targeting first paying customer, with both engineering milestones and business value framing.

Approach: We will generate this as a structured document (docs/plan/roadmap-v2.md) containing:

1. Complete feature inventory (what exists, what’s missing, what’s needed)
2. Prioritized milestone plan organized by releases
3. Business-value framing per milestone
4. Dependencies and critical path

---

Plan: Generate the Roadmap Document

Step 1: Create Feature Inventory (docs/plan/feature-inventory.md)

A comprehensive matrix of every feature and subfeature, classified by status. Organized into these 9 feature domains:

Domain 1: Core Design Pipeline

Feature	Subfeature	Status	Gap
Workload Analysis	Keyword-based estimation	DONE	Need learned estimation
Hardware Explorer	Static catalog + scoring	DONE	Need dynamic catalog from embodied-schemas
Architecture Composer	Operator→hardware mapping	DONE	Need multi-accelerator heterogeneous mapping
PPA Assessor	Power/latency/area/cost verdicts	DONE	Need calibrated models (±15% accuracy)
Critic Agent	Single-compute risk, power margin	DONE	Need expanded critique vocabulary
Report Generator	Structured reports + audit trail	DONE	Need PDF export, comparison reports
Optimization Loop	Framework defined (Phase 2 plan)	PARTIAL	Need full implementation: working memory, experience cache, strategy catalog
Planner + Dispatcher	LLM goal decomposition, DAG scheduler	DONE	Need streaming, retry policies

Domain 2: Co-Simulation & Validation (PHASE 1 — NOT STARTED)

Feature	Subfeature	Status	Gap
Simulator Integration	AirSim/Gazebo ↔ CSim bridge	NOT STARTED	Full implementation needed
Energy Traceability	Per-component power breakdown	NOT STARTED	Instrumentation layer needed
Latency Correlation	Sensor→actuator timing	NOT STARTED	Timestamp synchronization needed
Hardware Model Accuracy	±15% vs physical hardware	NOT STARTED	Empirical calibration needed
Dataset Generation	10K+ flight minutes, 50+ configs	NOT STARTED	Parallel simulation infra needed
Sim-to-Real Validation	Predicted vs measured comparison	NOT STARTED	Physical hardware test rig needed

Domain 3: Hardware Design & Synthesis

Feature	Subfeature	Status	Gap
KPU Micro-Architecture Config	3 presets, technology scaling	DONE	Need more presets, user-defined configs
Floorplan Validation	Checkerboard pitch matching	DONE	Need area-aware placement
Bandwidth Validation	4-link memory hierarchy check	DONE	Need NoC topology exploration
RTL Template Engine	12 Verilog modules via Jinja2	DONE	Need FPGA synthesis flow, timing closure
Manufacturing Cost Model	16 process nodes, yield model	DONE	Need packaging cost, volume discounts
EDA Tool Integration	Yosys/Verilator/Icarus pipeline	DONE	Need commercial EDA wrappers (Synopsys, Cadence)
IP Block Library	Basic MAC, SRAM, DMA, router	DONE	Need standardized IP catalog, licensing model

Domain 4: Benchmarking & Profiling

Feature	Subfeature	Status	Gap
Local CPU Backend	PyTorch inference timing	DONE	—
Local GPU Backend	CUDA benchmarking	NOT STARTED	Need GPU backend implementation
Remote SSH Backend	Framework + SFTP transfer	PARTIAL	Model serialization is stub
Kubernetes Backend	Job-based parallel benchmarking	PARTIAL	Model loading is stub
Power Monitoring	RAPL + AMD SMU backends	DONE	External power meters (SCPI/USB) not implemented
Architecture Benchmarking	Operator pipeline timing	DONE	—
Roofline Analysis	Via graphs package	DONE	—

Domain 5: Deployment Targets

Feature	Subfeature	Status	Gap
Jetson (TensorRT)	FP32/FP16/INT8 + validation	DONE	Need on-device profiling integration
Coral (EdgeTPU)	INT8/UINT8 + compiler	DONE	—
OpenVINO	FP32/FP16/INT8 + NNCF	DONE	—
Stillwater KPU	Full ISA + deployment	DONE	Need real silicon validation
NVDLA	Full spec + deployment	DONE	—
RISC-V targets	—	NOT STARTED	Need RISC-V toolchain integration
FPGA targets	—	NOT STARTED	Need Vivado/Quartus integration

Domain 6: CLI & Developer Experience

Feature	Subfeature	Status	Gap
19 CLI subcommands	Core commands registered	DONE	—
Backend CLI integration	SSH/K8s backend selection	STUB	CLI says “not yet implemented”, falls back to local
Backend management	branes backends add/test/configure	STUB	Must edit YAML manually
Config management	branes config set	STUB	Must edit YAML manually
Design synthesis	branes design synthesize	STUB	Synthesis engine placeholder
Spec template operations	Template variant wiring	PARTIAL	Some templates not fully wired
Interactive chat	Claude-powered agentic loop	DONE	—
Model Zoo	Multi-provider download/cache	DONE	—
Demo system	7 discoverable demos	DONE	—

Domain 7: Multi-Objective Optimization

Feature	Subfeature	Status	Gap
MAP-Elites	Quality-diversity search	DONE	—
Bayesian Optimization	BoTorch qNEHVI	DONE	predict() not implemented
NSGA-III fallback	Many-objective (>4)	DONE	—
MCP Server	5 optimization tools	DONE	—
CLI Integration	explore/show-front/sensitivity/explain	DONE	—
Surrogate Models	Fast approximation	NOT STARTED	Need trained surrogates for rapid iteration

Domain 8: Perception & Prototypes

Feature	Subfeature	Status	Gap
Drone Perception Pipeline	Detection + tracking + reasoning	DONE (Phase 3)	—
LiDAR Drivers	Velodyne/Ouster/Livox/ROS	STUB	5 drivers unimplemented
Stereo Fusion	Depth map integration	PARTIAL	Basic averaging only
Multi-Rate Framework	Zenoh pub/sub scheduler	DONE (prototype)	Need production hardening
Recording/Replay	HDF5 recording (Phase 4 planned)	NOT STARTED	—

Domain 9: Infrastructure & Quality

Feature	Subfeature	Status	Gap
CI/CD	Lint + test + release + docs deploy	DONE	Need integration test stage for optional deps
Documentation Site	Astro Starlight, 28 pages	DONE	Need API reference auto-generation
Test Suite	560+ tests, 47 test files	DONE	Need coverage reporting, integration tests in CI
Golden Traces	Regression detection	DONE	—
Cost Tracking	Token cost per agent	DONE	—
Safety Governance	Audit trail + flagging	DONE	Human approval is placeholder
MCP Protocol	Server implementation	DONE	—
Dataflow Inference	Kernel→kernel bandwidth	STUB	Hardcoded to 0 bytes, linear chains only

---

Step 2: Create Milestone Roadmap (docs/plan/roadmap-v2.md)

18-month timeline, 6 releases, targeting first paying customer.

Release 0.7 — “Production CLI” (Month 1-2)

Business Value: Tool is usable by early adopters without workarounds.

#	Feature	Priority	Effort
1	Wire SSH/K8s backends into CLI benchmark and workflow commands	P0	M
2	Implement branes config set with YAML validation	P0	S
3	Implement branes backends add/test/configure	P0	M
4	Fix Remote SSH model serialization (ONNX-based transfer)	P0	M
5	Fix Kubernetes model loading (ONNX from ConfigMap)	P0	M
6	Implement branes design synthesize (perception pipeline generation)	P1	L
7	Add Local GPU benchmark backend	P1	M
8	Bayesian MOO predict() implementation	P2	S
9	Wire all spec template operations	P2	S

Exit Criteria: User can branes benchmark run model.pt --backend remote_ssh and get results. Config/backends manageable via CLI.

---

Release 0.8 — “Optimization Loop” (Month 3-5)

Business Value: The agent autonomously improves designs — the core differentiator.

#	Feature	Priority	Effort
1	Implement Phase 2 optimization loop (working memory, strategy catalog)	P0	XL
2	Experience cache with SQLite + similarity search	P0	L
3	Governance: real human-in-the-loop approval blocking	P0	M
4	Dispatcher retry policies and streaming support	P1	M
5	Expanded critic vocabulary (thermal, reliability, supply chain)	P1	M
6	Surrogate model training for fast design space exploration	P1	L
7	Dataflow bandwidth inference (tensor shape → bytes)	P2	M
8	Complex dataflow topology (DAG support, not just linear chains)	P2	M

Exit Criteria: Agent takes a use-case description, iterates through ≥3 design alternatives, and converges on a design that meets all constraints with ≥20% improvement over naive baseline.

---

Release 0.9 — “Co-Simulation Foundation” (Month 6-9)

Business Value: Designs are validated against physics — customers can trust results.

#	Feature	Priority	Effort
1	AirSim/Gazebo ↔ compute simulator bridge	P0	XL
2	Per-component energy instrumentation layer	P0	L
3	End-to-end latency correlation (sensor → actuator)	P0	L
4	Hardware model calibration against physical boards (Jetson Orin)	P0	L
5	Mission profile library (hover, cruise, obstacle avoidance, etc.)	P1	M
6	Software configuration matrix (50+ configs)	P1	L
7	Parallel simulation infrastructure (10K flight minutes)	P1	L
8	External power monitoring (SCPI/USB meter integration)	P2	M

Exit Criteria: Run a simulated drone mission, get ±15% accurate energy/latency predictions vs physical Jetson Orin hardware.

---

Release 1.0 — “Customer-Ready Platform” (Month 10-12)

Business Value: First paying customer. Production-quality with documentation and support.

#	Feature	Priority	Effort
1	Report generation: PDF export, comparison reports	P0	M
2	API reference auto-generation from docstrings	P0	M
3	Integration tests in CI for all optional dependency groups	P0	L
4	RISC-V deployment target	P1	L
5	FPGA deployment target (Vivado/Quartus)	P1	XL
6	Multi-accelerator heterogeneous mapping in architecture composer	P1	L
7	Production multi-rate framework (from prototype to core)	P1	L
8	Test coverage reporting and 80%+ gate	P2	M
9	LiDAR driver implementations (Velodyne, Ouster)	P2	M

Exit Criteria: A customer can install pip install branes[all], analyze their codebase, get hardware recommendations, run benchmarks on remote hardware, validate in simulation, and receive a professional PDF report.

---

Release 1.1 — “Enterprise Features” (Month 13-15)

Business Value: Enterprise readiness — team collaboration, compliance, scale.

#	Feature	Priority	Effort
1	Team collaboration (shared specs, designs, reports)	P0	XL
2	Commercial EDA tool wrappers (Synopsys DC, Cadence Genus)	P0	L
3	IP block licensing model and catalog	P1	L
4	Safety certification compliance (IEC 62304, ISO 26262, DO-178C)	P1	XL
5	Proper stereo vision fusion (bilateral/guided filters)	P2	M
6	Recording/replay with HDF5 (drone perception Phase 4)	P2	M
7	Volume discount and packaging cost models	P2	S

Exit Criteria: Enterprise customer can use the tool in a regulated environment with audit trails, team sharing, and commercial EDA integration.

---

Release 1.2 — “Sim-to-Real Validation” (Month 16-18)

Business Value: Proof that the co-design thesis works — measurable improvement on real hardware.

#	Feature	Priority	Effort
1	Sim-to-real accuracy validation (±10% predicted vs measured)	P0	XL
2	Physical drone deployment and flight test	P0	XL
3	≥50% flight time improvement proof point	P0	L
4	Blueprint package (binary + HW config + BoM + report)	P0	M
5	NoC topology exploration in bandwidth validation	P1	L
6	Timing closure flow for RTL	P1	XL
7	LiDAR drivers: Livox, ROS bridge, file-based (.pcd/.ply)	P2	M
8	Learned workload estimation (replace keyword-based)	P2	L

Exit Criteria: Demonstrate a drone that flies 50% longer on co-designed HW/SW vs baseline, with predictions matching reality within 10%.

---

Step 3: Process for Generating These Documents

We will create two documents:

1. docs/plan/feature-inventory.md — The exhaustive feature matrix above, formatted as markdown tables with links to source files. This is the living inventory that gets updated as features are completed.

2. docs/plan/roadmap-v2.md — The milestone roadmap above, with:

   • Release timeline with dates
   • Per-release feature table with effort estimates (S/M/L/XL)
   • Business value statement per release
   • Exit criteria per release
   • Dependency graph showing critical path
   • Risk matrix per release

Step 4: Critical Path Analysis

Release 0.7 (CLI polish) ──┐
                            ├──→ Release 0.8 (Optimization loop) ──┐
                            │                                       │
                            └──→ Release 0.9 (Co-simulation) ──────┤
                                                                    │
                                                    Release 1.0 (Customer-ready) ──→ 1.1 ──→ 1.2

Critical path: 0.7 → 0.8 → 1.0 (the optimization loop is the core differentiator and must work before customer release)

Parallel path: 0.9 (co-simulation) can proceed in parallel with 0.8 if there are multiple engineers.

Files to Create/Modify

File	Action	Description
docs/plan/feature-inventory.md	CREATE	Exhaustive feature matrix with status and gaps
docs/plan/roadmap-v2.md	CREATE	18-month milestone roadmap with business framing

Verification

• Review both documents for completeness against the codebase exploration findings
• Ensure every TODO/STUB/gap from the codebase audit appears in the feature inventory
• Verify critical path makes sense given dependencies
• Cross-reference with existing docs/plan/roadmap.md to ensure nothing from the original plan is lost