Python Developer – Complete & Automate an Existing AI Pipeline (Claude-Assisted)

Remote | Philippines | Part-Time / Project-Based Initially | Long-Term Opportunity

We are looking for a technically strong Python / AI Automation Developer to help us finish and operate an existing conversation-intelligence platform.

This is not a general VA position, and it is not a greenfield software project.

The core Python application and AI evaluation engine are already built and working. We need someone who can step into the existing codebase, understand the architecture, use Claude/Claude Code effectively, complete the remaining integrations and automation, test/debug the system, and help us get it into production.

What We're Building

Callsistance analyzes recorded sales calls and automatically produces coaching intelligence for sales teams.

The completed pipeline will:

Receive call recording → Transcribe → AI evaluation → Generate PDFs → Create timestamped audio clips → Human QC/approval → Assemble manager/rep digests → Email delivery

The core transcription and multi-pass AI evaluation system is already functioning. The AI evaluation engine works with a structured configuration ("Blueprint") for each client — the specific scenarios, criteria, and scoring tiers it should evaluate calls against. Responses aren't unstructured AI output; they're generated and cited against that client's own defined structure.

The evaluation engine's current output format is considered stable, and the remaining work is intended to build on top of it rather than rebuild it. That said, if a later stage genuinely needs additional data that isn't currently included in that output (for example, more precise timing information for audio clips), a scoped adjustment to what gets passed through earlier in the pipeline is reasonable — that's a data/structure addition, not a rewrite of the evaluation logic itself.

The Milestones — Detailed

We've broken the remaining work into 9 milestones. We're giving more detail than usual below because we want quotes based on real understanding of the work, not guesses.

M1 — PDF Integration
Map the fields in our existing evaluation output (a structured JSON object per call — scores, categories, coaching notes, etc.) into our already-designed, locked HTML/CSS PDF template. The template and its rendering approach are fixed; this milestone is about building the adapter layer that feeds the template correctly, including chart generation for one data visualization on the report. Audio links and client logos are explicitly out of scope for this milestone (see M3 and M4) — they'll be stubbed.

M2 — Automated Report Generation
Wrap M1's rendering into an automated step that runs unattended as part of the pipeline: consistent file naming, output routing to the right storage location, and error handling so one failed report doesn't halt others.

M3 — Audio Clip Generation (Timestamped Clips + Listen Links)
The evaluation output references specific moments in a call (a quote or scored moment); this milestone cuts short audio clips (roughly 30 seconds) around those moments from the full call recording and produces working "listen" links embedded in the PDF. The central technical problem: reliably mapping an AI-cited moment back to the correct timestamp in the original audio, especially when the AI's phrasing doesn't match the transcript word-for-word. We're open to different approaches here (e.g., carrying word-level timing data through the pipeline vs. fuzzy-matching after the fact) and want to hear how you'd solve it before pricing it. Clips upload to cloud storage (Cloudflare R2) with a defined retention period.

M4 — Client Branding & Routing
Some clients get their own logo on reports and digest emails; others get our standard branding. This is config-driven per client, not a per-client code branch — a field on the client's config record determines which logo asset loads. No new evaluation logic, just asset routing.

M5 — Review / Approval Workflow
A queue where generated reports can be manually reviewed and approved before being included in a client's digest — this is the default mode for newer clients while we're still validating output quality. In parallel, this milestone needs a config-driven auto-approve mode (no human step at all) for clients where quality is already proven out — same underlying queue and data model, switched by a per-client setting, not two separate systems. State needs to survive a process restart (i.e., this can't be purely in-memory) since approval can happen at any point across a full day.

M6 — Salesperson-Level Reports & Digests
A variant of the manager report/digest, scoped to an individual salesperson's own calls only. This has a hard requirement: one rep must never be able to see another rep's data, enforced at the point data is queried/filtered — not just by what the UI or PDF happens to display.

M7 — Automated Intake
Pipeline currently kicks off from a manually placed audio file. This milestone builds the automated trigger — watching for new recordings and starting the pipeline without manual action. Needs to be safe against the same recording arriving twice (e.g., a retried upload) without spawning duplicate pipeline runs.

M8 — Manager + Salesperson Digest Delivery
Once-daily email digest per client, assembled from whatever reports are approved as of a fixed per-client send time. A report not yet approved at that moment is simply left out of that day's email and included automatically in the next day's once approved — never sent unapproved, never dropped. Needs to be safe against a retry sending the same digest twice.

M9 — Full Integration + Testing
End-to-end wiring of M1–M8 together, tested against real call recordings (not synthetic data), with a defined acceptance bar before we consider the initial build complete.

(A note on scope: some of the durability/state requirements above — e.g., "state must survive a restart," "must be safe against retries" — are real requirements, not boilerplate. Feel free to flag in your application if you'd want to confirm our current persistence approach before finalizing a number for M5, M7, or M8.)

VoIP API integration (Milestone 10) is scoped and priced separately, later, once the core pipeline is proven — see the note below. It is not included in this milestone list and is not required for this role.

What You'll Help Us Complete (quick-reference list, same scope as above)

Integrating existing AI evaluation JSON with our locked PDF templates
Automating PDF generation
Generating timestamped audio clips from call recordings
Uploading audio/report assets to cloud storage
Creating clickable audio links inside PDFs
Client-specific branding and white-label routing
Building a human report review/approval queue, including a config-driven auto-approve mode for proven clients
Creating manager daily digest emails
Creating salesperson-specific daily digests and coaching PDFs
Automating intake and processing
Deploying and monitoring the completed pipeline
End-to-end testing and debugging

A note on VoIP: further down the road, once the core pipeline is proven, we'll also need to connect a VoIP system through APIs/webhooks for automated call intake. This is not a requirement for this role. If telephony/webhook integrations aren't your area, that's completely fine — we're comfortable hiring a specialist for that piece separately once we're closer to that stage. If it is something you have real experience with, mention it in your application, but it will not affect your candidacy either way.

Current Technology

You do not need prior experience with every service.

Our current stack includes: Python, Anthropic/Claude, Deepgram, REST APIs and webhooks, Git/GitHub, Railway, Cloudflare R2, SendGrid, Google APIs, JSON, HTML/CSS PDF generation, VoIP APIs (see note above — not required).

Experience with FastAPI, Docker, cloud deployment, AI/LLM APIs, or automation systems is also valuable.

Claude Is Part of the Development Process

This is important:

We intentionally use Claude heavily to develop the application.

We are not looking for someone who believes every line of code needs to be written manually.

We also are not looking for someone who simply asks AI to write code and assumes the result is correct.

We want someone who can:

Give Claude strong technical instructions
Understand the code and architecture well enough to work with it
Run and test AI-generated code
Read logs and errors
Diagnose failures
Modify code directly when appropriate
Recognize when Claude's solution is wrong
Iterate until the required result actually works

If you already use Claude Code, Cursor, GitHub Copilot, ChatGPT, or similar tools for software development, we want to hear about it.

Who We're Looking For

You should be strong in: Python, API integration, REST APIs, Webhooks, Debugging, Git/GitHub, Automation, Working with an existing codebase, AI-assisted software development.

You should also:

Communicate clearly in English
Follow detailed specifications
Test your own work before submitting it
Document meaningful changes
Be comfortable explaining technical problems to non-programmer founders
Know when to ask a question rather than making an unapproved architecture change

We care more about demonstrated ability than degrees or certifications.

Project Structure & Compensation

We expect the initial build to be structured as fixed-price milestones rather than hourly billing, based on the 9 milestones detailed above.

As part of your application, we'd like you to propose:

A fixed price for each of the 9 milestones
Your best estimate of the total calendar time to go from Milestone 1 through Milestone 9 complete, given your stated availability

We understand both of these will be preliminary until you've seen the full technical scope and acceptance criteria, and we're glad to revise them together once you have it. If any milestone is genuinely hard to estimate without inspecting our existing code first, say so and tell us specifically what you'd want to see.

Long-Term Opportunity

We are not looking only for someone to finish a project and disappear.

If we work well together, we would like this person to potentially become our ongoing technical operations resource for the company.

Future work could include: pipeline monitoring, troubleshooting, client onboarding/configuration, new VoIP integrations, Blueprint/config updates, report QC support, API maintenance, automation improvements, continued Claude-assisted development, new product features.

The initial project is a good opportunity for both sides to determine whether there is a long-term fit.

HOW TO APPLY

Please answer all 7 questions below.

Generic applications that do not answer the questions will not be considered.

Tell us about the most relevant Python application or automation you've built. What did it do?
Describe an API/webhook integration you personally built or maintained. Which systems did you connect?
Which AI coding tools do you currently use (Claude Code, Claude, Cursor, Copilot, ChatGPT, etc.) and how do you use them during development?
Tell us about a time AI generated code that did not work correctly. How did you determine what was wrong and fix it?
What experience do you have deploying or maintaining Python applications in production? Which cloud/deployment platforms have you used?
Please provide a GitHub, portfolio, or example of relevant work if available.
After reading this listing, what part of this project do you believe would be the most technically challenging, and why?

Finally, please include:

Your normal availability
Your preferred working hours in Philippine Time
Your current hourly rate for reference, even though the initial project will primarily use fixed-price milestones
**Your proposed fixed price for each of the 9 milestones listed above**
**Your best estimate of total calendar time from Milestone 1 to Milestone 9 complete**

Important: Start your application with the word CANDLE so we know you read the entire listing.