Python Developer — Ongoing Role for AI Agent Fleet (Self-Correcting Agents + Landing-Page Quality) -- 3

We run a fleet of autonomous AI agents (Python, FastAPI, Celery, PostgreSQL, Docker) operating
roughly 10 small-business platforms in Saudi Arabia (e-commerce store builder, logistics/booking,
tutoring/education, personal & business finance tracking, real estate, and a few more), aimed at
real paying customers, not a demo. We recently ran a small paid test task with a developer. It
went well — real findings, independently verified — and it surfaced exactly what the real,
ongoing work actually is. This posting describes that work in full, openly, so anyone qualified
can propose their own approach and price. It is not a continuation reserved for one person.

**The honest current state (so you know what you're bidding on, not a sales pitch):**
- Our flagship platform (automated merchant landing-page builder) has 1,559 merchant records in
the pipeline. Exactly 1 page has ever been published live, and it was a free gift to a family
member, not a sale. ~1,530 records are stuck one step before any revenue-generating action.
- We had an 8-day period where every agent stopped doing real work while our own automated health
checks kept reporting "healthy" the entire time, because those checks never actually exercised
the AI layer — only a human manually noticing ended it. Likely cause: the whole fleet shares one
subscription-based inference pool with a weekly usage ceiling; we don't yet have proper usage
telemetry to prove it, and that's part of the ask below.
- An internal technical audit (self-run by the agent fleet, ~34 sub-investigations, reviewed by a
human) found real, unglamorous technical debt across the fleet: duplicated/redundant scheduled
jobs burning API budget for no functional gain, internal documentation that disagrees with what
is actually deployed and running (in several places), the same third-party integration installed
as multiple diverging copies instead of one shared version, at least one background job silently
failing to persist real customer submissions with no alerting, and general drift between what
the docs say is live/frozen and what's actually running. None of this is exotic — it's the
normal accumulated debt of a fast-moving small operation, and we'd rather describe it honestly
than pretend the fleet is pristine.
- Token/cost accounting exists but is only partially trustworthy — a real total is trackable, but
a meaningful chunk of individual run records don't carry usable cost data, so anyone doing
cost-optimization work here needs to fix the measurement before trusting it.

**Two concrete, currently-unstaffed work items (this is the actual core of the role):**

**1) Agents that actually learn from how we work with them.**
Our agents do not currently read their own conversation history with the business owner to
extract how he thinks, what he repeats, and how he follows up — and feed that back into
correcting their own automated workflows ("loops") and pipelines. Today that correction is
manual and inconsistent, entirely dependent on a human noticing and re-explaining the same thing
repeatedly. We need someone to design and build a real mechanism: what session/conversation data
to use, how to extract durable patterns vs. noise, what confidence bar before a correction is
applied automatically vs. flagged for review, and what it should actually feed into (loop
configs, prompts, escalation rules, standing decisions). This is acknowledged up front as
product-scale build work, not a quick script — come with a design proposal, not just an hourly
rate.

**2) Landing pages that actually sell.**
The store-builder platform above auto-generates merchant landing pages. They render correctly and
pass technical QA, but they don't clear a "sells itself to a real visitor in the first few
seconds" bar — a previous contractor also attempted and did not fix this. We need someone who can
actually diagnose and fix the generation pipeline (content strategy, structural layout, or model
prompting — your call, defend your approach in your proposal) against a real, already-existing
internal quality gate, proven on a real live store with a genuine before/after, not a mockup.

**Also realistically in scope over time, lower priority than 1 & 2 but real:**
cost/token telemetry that can actually be trusted, de-duplicating redundant scheduled automation,
reconciling documentation against live reality on an ongoing basis, and generally raising the
engineering hygiene bar across ~10 projects run by a fleet of AI agents with light human
oversight. We'd rather a good hire absorb this over time than pretend it isn't there.

**What is intentionally NOT in this public posting, and why:** exact infrastructure topology,
credentials, internal network/container details, and specific unpatched technical issues. This is
a public listing anyone can read, including people with no relationship to us — publishing
exploitable specifics here would be actively irresponsible, not "hiding scope." Nothing about the
SIZE, NATURE, or SEVERITY of the work above is being softened or hidden; a detailed internal
technical report exists and will be shared privately with serious, shortlisted candidates once
we're discussing real pricing.

Real programming experience required: Python, REST/OAuth APIs, task queues (Celery or similar),
PostgreSQL, Docker. LLM/agent-tooling experience (prompting, evals, RAG, tool-calling agents) is a
strong plus for item 1 specifically. Comfort reading and improving someone else's live production
codebase is required — this is not a greenfield project. Remote, any location. Clear written
English required, since day-to-day communication and written proposals will be in English even
though the end product serves Arabic-speaking customers.

We will start with a well-defined, paid opening phase on one of the two core items above — your
choice which, argue for it in your proposal — then move to an ongoing arrangement (hourly or
milestone-based, your preference) if it's a good fit on both sides. Every proposal will be
reviewed on its actual approach and reasoning, not just the number attached to it.