Building JuntoAI Echo in 8 Days: 14 POCs, 144 Commits, and a Lot of AI Tokens
How I went from zero to a production AI copywriter in eight days using spec-driven development and four concurrent AI agents.
Building the future of professional networking with AI | AWS & Data Architect | CTO @ JuntoAI
On March 7, 2026, the first commit for JuntoAI Echo landed. Eight days later, on March 15, we had a production AI copywriter that interviews you for 8 minutes, then writes LinkedIn posts that sound like you. Complete with image generation, LinkedIn scheduling, an admin dashboard, content transparency scoring, and GDPR compliance specs.
This is the story of how that happened. Not the polished version :-)
The Starting Point
Echo wasn't built from scratch. It was built on top of JuntoAI Kinetic , our career interview platform that was already running in production with Amazon Nova 2 Sonic. We'd spent weeks building Kinetic's voice infrastructure: the Bedrock Sonic wrapper, WebSocket utilities, stream management, session lifecycle, and structured logging. All of that lived in a shared library called
juntoai-core .The hypothesis was simple: if we could interview someone about their career in 15 minutes, we could interview them about their week in 8 minutes and turn it into LinkedIn posts. Same voice AI. Different persona. Different output.
But "different persona, different output" turned out to be 14 proof-of-concept specs, 144 git commits, and a complete product.
The Eight Days
Day 1 (March 7): Foundation
2 commits. POC-1 landed: the copywriter backend service, shared library integration, frontend scaffold, and start scripts. The spec had been written the day before. By end of day, we had a working voice interview with a copywriter persona on port 3005.
Day 2 (March 8): The Big Push
36 commits. The most intense day. POC-2 (post generation pipeline: style analyzer, post generator, image generator, calendar scheduler, quality validator) and POC-3 (LinkedIn scheduling loop: post editor, scheduler, LinkedIn client, email client, content manager, cycle history) both shipped. The frontend got its landing page, branding, trust bar, and release notes. We hit 723 passing tests. The CI/CD pipeline broke three times on Docker path issues and lock file sync problems.
Day 3 (March 9): Interview Intelligence
10 commits. POC-6 added a background transcript evaluator that watches the interview in real time, detecting when enough post-worthy topics have been covered. Sliding-window evaluation with topic carry-forward across evaluator calls.
Day 4 (March 10): The Rewrite
11 commits. POC-7 replaced the entire interview architecture. The 15-minute multi-stream rotation system became a single 8-minute continuous stream. Nova 2 Sonic's native tool use replaced the external topic steering. Split-tier evaluation (fast counts every minute, deep analysis at T=4 and T=7). A ghostwriter pipeline with insight extraction, voice imprint, angle selection, and hook generation. This was the biggest architectural bet of the project.
Day 5 (March 11): Stabilization
5 commits. The quietest day. Fixing edge cases from the POC-7 rewrite. Stream failure recovery. Silence handling. Early end confirmation.
Day 6 (March 12): Scheduling & Admin
28 commits. POC-8 (cloud-native scheduling via EventBridge), POC-9 (LinkedIn post management), POC-10 (auto-generated attribution comments), and POC-11 (admin dashboard with JWT auth, user listing, session detail, pipeline monitoring). Four POCs in one day.
Day 7 (March 13): Image & Intelligence
9 commits. POC-12 switched image generation from DALL-E 3 to Google Imagen 4 as the primary provider. POC-13 added auto-generated hashtags. POC-14 introduced content transparency: per-topic readiness signals, information density metrics, and Content DNA labels showing how much of each post is the user's words versus AI structure.
Day 8 (March 14): Release
38 commits. v0.2.0 released. Final bug fixes, test fixes, the Imagen 4 PR merged, content transparency wired end-to-end. The release notes page listed 30+ features across 7 POCs.
The Spec Machine
Every one of those 14 POCs followed the same three-document workflow: requirements, design, tasks. The requirements used SHALL-style acceptance criteria. The design included architecture decisions with alternatives considered. The tasks referenced specific requirements and file paths.
The specs weren't documentation written after the fact. They were the implementation instructions. The AI agents read the spec, understood the context from steering files, and executed the tasks. The human reviewed at checkpoints.
Here's the POC sequence and what each one delivered:
POC | What It Built |
|---|---|
POC-1 | Voice interview engine, onboarding, persona, infrastructure, landing page, auth, shared library |
POC-2 | Style analysis, post generation (2-5 posts), image generation, content calendar, quality validation |
POC-3 | Post editor, LinkedIn scheduling, weekly email loop, attribution comments, cycle management |
POC-6 | Background transcript evaluator, topic detection, finish condition, auto wrap-up |
POC-7 | 8-minute single-stream rewrite, native tool use, ghostwriter pipeline, voice imprint |
POC-8 | EventBridge scheduler, Lambda publishing, post-published notifications |
POC-9 | View on LinkedIn, copy URL, delete published posts |
POC-10 | AI-generated attribution comments, CTA comments, per-post toggles |
POC-11 | Admin dashboard, user listing, session detail, pipeline monitoring, global search |
POC-12 | Imagen 4 as primary image provider, context-aware prompts, extended image types |
POC-13 | Auto-generated hashtags (2-4 per post), validation, deduplication |
POC-14 | Per-topic readiness signals, information density, Content DNA labels |
Plus POC-20 through POC-25 were specced (multilanguage, analytics, personal photos, advanced settings, WhatsApp interviews, GDPR compliance) for the next release cycle.
The Numbers
Development span: 8 days (March 7-14, 2026)
Git commits: 144 (across backend and frontend)
POCs implemented: 14 (v0.1: 5 POCs, v0.2: 7 POCs, plus 2 architectural rewrites)
POCs specced for future: 6 more (POC-20 through POC-25)
Spec documents written: 42+ (3 per POC: requirements, design, tasks)
Peak commit day: 38 commits (March 14, release day)
Busiest build day: 36 commits (March 8, POC-2 + POC-3)
Codebase:
- Backend service: ~30 source files across services, routes, middleware, config
- Frontend app: ~25 components, pages, hooks, and utilities
- Shared library: 5 core modules (Bedrock Sonic, stream manager, WebSocket utils, logger, session base)
- Tests: Hundreds of unit and property tests across backend, frontend, and shared library
- Infrastructure: 10 Terraform modules (ECS, DynamoDB, S3, ALB, CodePipeline, CloudFront, CloudWatch, SSM, ECR, Lambda)
AI models in production:
- Amazon Nova 2 Sonic: Real-time speech-to-speech interviews
- Amazon Nova 2 Lite: Style analysis, post generation, content signal extraction, CV parsing, topic evaluation
- Google Imagen 4: Post image generation
- Cost per weekly cycle: ~€0.33 (interview + analysis + posts + images)
What Made This Possible
Three things:
1. The shared library.
juntoai-core gave us the entire voice infrastructure for free. Bedrock Sonic wrapper, WebSocket handling, stream management, session lifecycle, structured logging. Echo imported it as an npm dependency and customized via configuration. No code duplication. The hardest engineering (Nova Sonic's event protocol, stream rotation, transcript filtering) was already battle-tested in Kinetic.2. Spec-driven development.
Every POC had a spec before a single line of code was written. The AI agents didn't guess what to build. They read requirements with acceptance criteria, design documents with architecture decisions, and task lists with file paths. The specs were the bridge between human intent and machine execution.
3. Concurrent AI agents.
Four AI agents running simultaneously across two machines. One on infrastructure. One on backend. One on frontend. One on tests. The human coordinated, reviewed, and made architectural decisions. The agents executed. This is the setup described in blog post 15 of this series, and Echo was the project that proved it works at scale.
The Honest Parts
It wasn't all smooth. The CI/CD pipeline broke repeatedly on lock file sync issues (we've since added a hook to prevent that). The POC-7 rewrite was a gamble that could have cost two days if it failed. Some days had more bug-fix commits than feature commits. The admin dashboard shipped with a login rate limiter because we forgot to add auth on the first pass.
And the token consumption was significant. Running four concurrent AI agents for eight days burns through AI credits fast. We hit our Kiro Pro+ allocation limit and had to manage the workflow to degrade gracefully (that's the story in blog post next week :-) ).
But the product shipped. It's live at echo.juntoai.org . Real users do real interviews and get real LinkedIn posts that sound like them.
What This Means
Eight days from zero to production for a complete AI product isn't normal. It's what happens when you combine:
- A shared infrastructure library that eliminates the hardest engineering
- A spec workflow that gives AI agents precise implementation context
- Multiple concurrent agents that parallelize the work
- A human who makes decisions and stays out of the execution path
The AI didn't build Echo. A human designed Echo, wrote the specs, made every architectural decision, and reviewed every checkpoint. The AI executed the implementation at a speed that would be impossible for a human team of any size.
That's the model.
Not AI replacing developers.
AI multiplying one developer's decisions into a shipping product.
JuntoAI Echo is live at echo.juntoai.org . Built in 8 days with Amazon Nova 2 Sonic, Nova 2 Lite, and Google Imagen 4. Try it free, an 8-minute voice interview is all it takes.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article