Allow Google Analytics to measure how I use the site to enhance functionality and service.
Allows us to show personalized advertisements based on your browsing behavior.
Share my activity data with cookies for targeted advertising.
Customize my ad experience by allowing cookies to personalize the ads I see.
Get clarity on what to fix, keep, and improve.
A small local-government software vendor had used Replit Agent to build a municipal meeting minutes app. Authentication, dashboards, user management, and meeting uploads already worked, and the platform was set up for an initial group of municipalities. What the client lacked was an independent engineering answer. Could this codebase carry production use, more municipalities, and the AI workflow planned for Phase 1, or did it need a rebuild first? They brought in TechAvidus to answer that before spending more, and to build the AI workflow itself.
Authentication, dashboards, user management, and meeting uploads.
Could the codebase support production use, more municipalities, and the planned AI workflow?
Assess the existing codebase independently and build the Phase 1 AI workflow before deciding whether a rebuild is necessary.
Clerks were still producing minutes by hand: transcribe the recording, format it, review it, publish it. A two-hour meeting could take 3–4 hours of admin work. The client wanted to automate municipality meeting minutes and add more municipalities, but couldn't commit budget to either until someone confirmed the Replit Agent foundation would hold. A wrong call cost money in both directions: rebuilding code that didn't need it, or scaling code that couldn't take it.
A two-hour meeting could require 3–4 hours of administrative work.
The only record of how the system worked was the codebase itself. Replit Agent had generated it, and no separate documentation existed.
The data model was multi-tenant, with logical isolation between municipalities. In Phase 1, municipalities were provisioned directly in the database rather than through an admin interface, so isolation had to be confirmed in the code, not assumed.
Speech-to-text output, agenda alignment, and the generated minutes all had to stay consistent across different recording conditions, speakers, and meeting formats.
Agendas arrive as Word or PDF documents, and each transcript had to be matched to the right agenda items before any minutes were drafted.
Every AI-generated draft had to stay editable and pass a clerk's review before it could become a permanent public record.
The client didn't need to build an AI meeting minutes app from scratch. It needed the AI workflow added to the one it already had, once a team had confirmed that app was worth building on. Over two months in 2025, a three-person TechAvidus team delivered:
From codebase assessment to AI workflow and future planning
A repository audit of the Replit Agent-generated application against the approved Phase 1 Scope of Work, with an architecture review and a walkthrough of the municipality workflow as implemented in code.
The minor improvements the audit identified in the existing code.
An OpenAI Whisper speech-to-text pipeline, plus an LLM-driven workflow that generates draft meeting minutes from each transcript and its uploaded agenda.
A scalability review for adding more municipalities.
An assessment of AI operating costs, with optimization work scheduled for a later phase.
An engineering roadmap for Phase 2, walked through with the client's stakeholders in review sessions.
A self-serve interface for creating municipalities and expanded reporting stayed on the Phase 2 roadmap. Everything else in the Phase 1 scope was validated and delivered.
The roadmap only made sense once the team knew what it was building on, so the work started with a production readiness assessment, not a rewrite estimate.
With no documentation to work from, the team read the Phase 1 Scope of Work and audited the repository directly, comparing what was implemented against what had been contracted. That gave an architecture review of an AI-generated app grounded in the client's own scope rather than in general opinions about AI-generated code. Findings were checked with the client's stakeholders in joint review sessions.
Authentication, dashboards, user management, and meeting uploads were working as intended and needed only minor improvements. User management, municipality administration, meeting processing, and the AI workflow already sat in separate areas of the code. A full rewrite was the alternative, and it was rejected: keeping the stable components meant less delivery risk, less effort, and lower cost than redeveloping them. New engineering went into municipality-specific logic, the AI workflow, and production readiness.
Transcription and minutes generation serve one objective, turning a meeting into a draft record. So the AI steps live in the meeting-processing workflow, not spread across the rest of the app.
Building Phase 2 features on top of unchecked code would compound whatever technical debt it held. Validation was completed first. AI token usage was reviewed during production-readiness planning, and the cost-optimization work itself was scheduled for after Phase 2.
The preserved application and the new AI workflow now run as one system, organized in four layers:
A single orchestration layer controls the order of the AI steps for each meeting:
Tenant isolation runs underneath all of it. Every meeting belongs to exactly one municipality, and a user's municipality is applied at login and held for the whole session. That means the same platform can take on more municipalities without one municipality's recordings or minutes becoming visible to another.
A municipality user creates a meeting and uploads an MP3 or WAV recording and the agenda.
The system checks the file formats, confirms an agenda is attached, links both files to the user's municipality, and blocks duplicate meetings.
OpenAI Whisper transcribes the recording.
The transcript is aligned to the items in the uploaded agenda.
An LLM generates structured draft minutes from the aligned transcript.
The draft appears in the review interface, where a clerk edits and approves it.
Once published, the minutes are stored as a permanent record and can no longer be edited.
TechAvidus stabilized the existing React and Supabase app, wrapped for Android and iOS with Capacitor and Cordova.
Staff create a meeting, attach the recording and agenda, and the system queues it for processing. This step was preserved and verified against the Phase 1 scope. It matters because the AI pipeline is only as dependable as what it receives: unsupported audio formats, missing agendas, and duplicate meetings are stopped here.
Whisper turns the recording into a transcript, the transcript is matched to the agenda, and an LLM writes structured draft minutes. The hard part is consistency across recordings with different speakers, audio quality, and meeting formats. For a clerk, it means starting from a first draft instead of a blank page.
Clerks correct and save drafts as many times as they need. Only authorized users can publish, and published minutes become immutable. A person stays responsible for every official record, and no AI output reaches the public record without that review.
Past meetings and their minutes can be searched, filtered, and reopened, always limited to the user's own municipality. Staff can pull up an earlier meeting from the platform itself instead of keeping separate records.
Municipalities, users, and roles are set up with tenant-isolated, role-based access. In Phase 1, municipalities are created directly in the database. A self-serve admin interface is on the Phase 2 roadmap.
Four integrations make up the AI workflow, all added to the preserved application rather than built as a separate system:
From raw meeting audio to a review-ready draft
Converts uploaded meeting recordings into machine-readable transcripts
Generates structured draft minutes from each transcript together with its agenda
Ingests Word and PDF agendas before minutes are generated
Sequences transcription, agenda alignment, generation, and the hand-off to review
The difficult part was alignment. The output had to be a structured draft that a non-technical clerk can edit with confidence, built from transcripts of very different recordings and matched to agendas formatted however each municipality writes them.
From an existing application to an AI-powered meeting minutes workflow.
Municipality staff can now turn an uploaded recording and agenda into an editable draft of the meeting minutes, using a Whisper speech-to-text pipeline and an LLM minutes-generation workflow, both built in this engagement.
No rebuild was needed. The independent audit confirmed that the Replit Agent-generated authentication, dashboards, user management, and meeting uploads were solid enough to keep, and the minor improvements it identified were applied.
The AI workflow was built inside the existing React.js, Node.js, and PostgreSQL codebase, without replacing the underlying stack.
The full approved Phase 1 Scope of Work was validated and delivered in 2 months.
A scalability assessment, an AI operating-cost assessment, and a Phase 2 engineering roadmap were delivered alongside the build.
If you're trying to fix a Replit-generated app without rebuilding it, here is what this project demonstrates.
The audit measured an undocumented Replit Agent codebase against its Phase 1 contract and found the foundation worth keeping. The client avoided paying for a rewrite it didn't need. This is how TechAvidus approaches Vibe Coding Rescue.
For a team that wants to develop an AI meeting minutes platform for municipalities on the code it already has, this shows a Whisper transcription and LLM drafting workflow built into an existing, constrained codebase, with a human review step before anything is published. See AI integration into existing software.
The client finished Phase 1 with a scalability assessment, an AI cost assessment, and a Phase 2 plan, so the next budget decision rests on engineering findings instead of guesswork.
