Turn every GitHub release into approved, on-brand, multi-channel launch content — with claim-level provenance, human approval at every step, and a system that learns your brand voice over time.
Built for the AWS hackathon · h01.devpost.com. Live on Vercel, with Amazon Aurora PostgreSQL as the system of record.
Engineering ships releases constantly. Turning each one into the blog post, changelog, social posts, customer email, and demo video is slow, manual, and off-brand — marketing re-writes what engineering already wrote, deadlines slip, the voice drifts, and no one can trace a marketing claim back to the code that justifies it. It's expensive, repetitive work that every software company does and nobody enjoys.
ShipSignal reads a release's GitHub diff and produces the full launch kit — release blog, changelog, LinkedIn/X posts, customer email, and a narrated audio + video digest — under three guarantees that make it safe to actually use:
- Evidence-backed, never hallucinated. It never writes copy from a raw diff. It first builds an evidence-backed feature manifest, and every generated claim links to the concrete code/PR/issue evidence it came from — viewable as a provenance graph.
- Human-governed. Three mandatory approval gates — feature manifest → generated artifacts → skill change — plus Amazon Bedrock Guardrails and deterministic PII/secret redaction. Nothing publishes without a human.
- Self-improving. Reviewer edits and rejections are mined into a proposed next-version skill; once a human approves it, the system writes better, more on-brand content on the next release. It compounds.
The payoff: diff → governed, on-brand, traceable, multi-channel launch content — and one-click distribution (e.g. publish the demo video to YouTube).
Software / dev-tools companies, DevRel, and product-marketing teams that ship often and need launch content that is fast, on-brand, traceable, and governed — without the manual grind.
A GitHub diff flows through a LangGraph pipeline, with humans in the loop at three gates and a learning loop feeding back in:
flowchart LR
A["GitHub release diff"] --> B["Evidence + signals<br/>redacted, in Aurora"]
B --> C{"Gate #1<br/>feature manifest"}
C -->|approved| D["On-brand content<br/>blog, changelog, social, email"]
D --> E{"Gate #2<br/>artifacts + provenance"}
E -->|approved| F["Narrated media<br/>audio + video, S3"]
D -. reviewer edits .-> G{"Gate #3<br/>skill evolution"}
G -.->|approved| D
- Two detailed diagrams — full system architecture and a dedicated skill-evolution (Gate #3) flow — are in demo/ARCHITECTURE.md.
Every entity lives in Aurora; the schema encodes the product's guarantees across 38 migrations:
- Tenancy by construction — every row is
release_run_id-scoped with cascading FKs, so a GDPR erasure is a single delete. - A provenance graph —
artifact_claims → evidence links → evidence_items; no unlinkable claim is stored approved. - Real
pgvectorsemantic search —vector(1536)+ HNSW, populated with real Amazon Bedrock Titan embeddings (verified cosine retrieval). - Behaviour-as-data — versioned skills, capability/agent governance maps, and a self-learning ledger let the system evolve without code changes.
- Encrypted connections — OAuth tokens stored AES-256-GCM-encrypted for one-click publishing.
Hackathon alignment (h01.devpost.com)
| Requirement / criterion | How ShipSignal meets it |
|---|---|
| AWS database | Amazon Aurora PostgreSQL (Serverless v2) + pgvector — the single source of truth (38 migrations, provenance graph, real vector search) |
| Deployed front-end (Vercel) | Next.js/React 19 dashboard live at shipsignal-xi.vercel.app, reading Aurora over verified TLS |
| Technological implementation / DB | Deep, deliberate schema; real ingestion → persistence; real Bedrock Nova authoring + Titan embeddings |
| Design | Three-gate review/approval UX, evidence→claim provenance views, media preview + 1-click publish |
| Impact | A real, recurring, expensive workflow (every release) turned governed and traceable |
| Originality | Not "diff → blogpost" — an evidence-backed manifest → human-gated, claim-traceable content → a self-learning skill loop |
Two end-to-end runs are live on the dashboard:
- Primary —
OrcaQubits/agentic-commerce-skills-plugins(run3b1fed7f): fully real — real diff → ~8,100 evidence rows → real Titan embeddings → real Amazon Bedrock Nova authoring → real ElevenLabs/ffmpeg media. - Secondary —
NousResearch/hermes-agent(run49a31f1c): the same, with the LLM authoring on an offline model (a deterministic, media-rich comparison).
Full honesty matrix (real vs. demo, with verification) → demo/VALIDATION.md.
| Artifact | What it is |
|---|---|
| < 3-min walkthrough of the live app (the agentic-commerce run, end to end) | |
| demo/SUBMISSION.md | The Devpost write-up — criteria-mapped, with live links + Team ID |
| demo/ARCHITECTURE.md | Architecture + skill-evolution diagrams, and the Aurora data-model deep dive |
| demo/DEMO_SCRIPT.md | The < 3-minute demo-video script (beat by beat) |
| demo/VALIDATION.md | Validation matrix — what's real, how it was verified |
| demo/assets/agentic_commerce_digest.mp3 · .mp4 | Real narrated media for the primary run (ElevenLabs + ffmpeg) |
Frontend/API: Next.js (App Router) + React 19 on Vercel ·
Database: Amazon Aurora PostgreSQL (Serverless v2) + pgvector ·
Storage: Amazon S3 (presigned-only media) ·
LLM/AI: Amazon Bedrock — Nova (authoring) + Titan (embeddings) + Guardrails ·
Orchestration: LangGraph (Python worker, 4 graphs) ·
Media: ElevenLabs TTS + ffmpeg.
38 Aurora migrations · 485 TypeScript + 420 Python tests green · deployed and live · two real end-to-end runs.
Licensed under the Apache License 2.0. See NOTICE for attribution.