01·Customer-facing AI at super-app scale
AI support, in production
As the company's first AI product manager, I shipped a leading delivery super-app's first customer-facing AI product: instant, high-quality support across Lebanon, Iraq, and Kurdistan, in English, Arabic, Kurdish Sorani, Kurdish Badini, and Turkish.
The problem
At super-app scale, support queues fill with routine questions, and every routine chat in the queue slows down the customer with a genuinely hard problem. The answer isn’t more agents; it’s making the routine instant and the path to a human faster.
It reshaped the support experience on both sides: no queue for the routine, and a faster path to a person for everything else.
How I shipped it
Trust is earned, not assumed. Gated, staged rollouts; evaluation frameworks; and the support team held full override control from day one.
Evals are the new requirements. I wrote scripted conversations with an expected outcome for every scenario, then ran and scored each one, conversation by conversation. Less time writing requirements, more time defining what a good answer looks like.
Behavior is the product. Updating the system no longer requires code or traditional requirements documents; it requires clear specifications and high-quality evals.
1st
customer-facing AI product at the company
0
languages: English, Arabic, two Kurdish variants, Turkish
0
markets: Lebanon, Iraq, Kurdistan
0
queue for the routine questions
“Text responses are just step one. Next is autonomous execution: canceling orders, updating delivery addresses mid-route, resolving delays before a customer even notices.”
02·Measuring the new search
GEO Monitor
When customers ask ChatGPT instead of Google, does your brand come up? GEO Monitor submits the question a customer would actually type to three AI engines and turns their answers into a score you can act on.
The problem
AI assistants increasingly answer “who should I hire?” directly, and brands have no visibility into whether they are recommended, buried, or absent, nor which sources the engines are citing instead of them.
What I built
The same brand question goes to ChatGPT, Gemini, and Claude with web grounding on. A fourth model pass reads all three answers and extracts structure: was the brand mentioned, at what rank, described how, on the strength of which sources. Scores are discounted by rank and sentiment into a weighted visibility metric, tracked sweep over sweep.
Engines do not give the same answer twice, so the design accounts for it: repeated sampling with honest variance reporting, rather than pretending one answer is the truth.
3
AI engines queried with web grounding
4th
model pass extracts structured findings
2
scores: raw visibility and rank-weighted
0
framework dependencies in the core engine
Brand dashboard: visibility climbing sweep over sweep as improvements land
Query by query: which engines mentioned the brand, at what rank, and why it was absent
The demo shows a seeded fictional brand; no client data appears anywhere.
03·Solar intelligence, before the sale
Solas
Solas turns raw generator data into a bankable proposal: it measures the site's real load, models the array against live irradiation data, and produces a one-page trilingual investment study you can hand across a desk.
The problem
Most commercial sites in Lebanon run on diesel generators, and solar pitches are mostly guesswork; nobody measures the actual load before quoting. Buyers are asked to trust a brochure.
What I built
Solas reads a generator-controller export, reconstructs the site’s load profile, pulls live irradiation data for the exact coordinates (PVGIS, EU Joint Research Centre), and produces a one-page trilingual investment study covering headline payback, system sizing, and battery opportunity, delivered as a PDF.
62
automated tests across the Solas–Solis pair
2
data sources joined: controller export and live irradiation
1
page: the whole study, ready to hand over
3
languages: English, French, Arabic with full RTL
Solas: live irradiation data, array sizing, and the payback headline
The demo loads a sample site in one click, or download the raw controller export .xlsx and upload it like a real one. The study above was produced from exactly that flow. Download the one-page PDF it generates →
04·Solar intelligence, after the installation
Solis
Commissioning day is where accountability usually ends. Solis reads the array's monitoring exports, rebuilds what the system should have produced for that exact location and period, and says in plain language whether it did.
The problem
Nobody verifies the array after commissioning. The proposal promised a number, the roof produces another, and a weak or shaded panel string can hide inside monthly totals for years.
What I built
Solis reads monitoring exports from six inverter brands with zero configuration, rebuilds the expected output for that location and period, and issues a plain-language verdict, including per-string health, so a shaded or failing panel section is caught from the data alone.
6
inverter vendors auto-detected from the file itself
0
configuration needed before the verdict
24
adversarial data scenarios in the parser suite
1
weak panel string is enough to get flagged
Solis: plain-language verdicts identifying a weak panel string
The demo loads a sample array in one click, or download the raw inverter export .csv and upload it like a real one.
05·Platform product management at scale
Toters
The AI product above is one of several programs I drive as a Platform Product Manager at Toters, the delivery super-app serving Lebanon and Iraq. Two further programs follow, described at case-study level: production work across customer, courier, and back-office applications.
Address revamp
Reducing undelivered orders
In markets where street addresses are often approximate, vague addresses fail deliveries and flood support. I own the program rebuilding addressing end to end: a richer address model with landmarks, photos, and voice notes; interception of incomplete addresses at checkout; and ops tooling to flag problem addresses back to the customer. Success is measured where it matters: fewer “can’t find the customer” escalations and faster driver time-to-find.
Platform integrity
Device intelligence with Incognia
I drive Toters’ fraud-prevention program: a fraud engine built behind a provider abstraction, rolled out monitor-first behind feature flags, powering device-level blocklisting, promo-abuse prevention, and SMS-pumping defense. Represented Toters at the closed-door Platform Integrity Network workshop in Singapore, co-presenting our approach alongside our partner Incognia.
About
I’m Charles. I work at the seam between product and engineering: close enough to the business to know which number matters, close enough to the code to ship it.
Beyond the case studies above, I’ve built and shipped software for restaurants, insurance agencies, and hospitality operations, in English, French, and Arabic.
Walk the farm
The portfolio, playable. Every landmark represents a product that shipped, and the full write-ups are readable from inside.
You’re standing on the portfolio.
click to take the view · WASD to move · space to jump · shift to crouch
right click or E interacts · left click swings the hoe
on touch: hold to walk, drag to turn, tap to interact
the farmer by the wheat has work for you · prefer reading? scroll back up ↑