.png&w=256&q=75)
Syed Aoun Haider Naqvi@noneed109888275
1
Events attended
1
Submissions made
Pakistan
1 year of experience
About me
I’m Aoun, and I’ve always been curious about how things work. That curiosity pulled me into AI and machine learning, and I taught myself by building things, breaking them, and figuring out why. I love exploring new tech, and I get a real kick out of turning a rough idea into something that actually runs. I’m here to learn from other builders, try ideas I haven’t tried before, and create something useful.
🤓 Latest Submissions
LedgerPilot AI
Accounts-payable teams still match invoices against purchase orders and delivery notes by hand. Two kinds of loss get through: overbilling (a higher price than agreed, units that never arrived, totals that don't add up), and payment diversion, where a perfect-looking invoice asks for payment to a new bank account or arrives twice. LedgerPilot does the matching and stops those payments before they go out. Evolus document extractors read the PO, the goods-received note and the invoice. Plain Python validates every field and runs the checks with Decimal maths: a three-way match on price and quantity, arithmetic and tax, the vendor master, bank details, PO balance, and exact and look-alike duplicates. An Evolus agent, Detective Clerk, whose model is the open-weight DeepSeek-V4-Flash served on AMD GPUs through AMD Token Factory, explains the evidence, cites the vendor master and proposes a route. An Evolus AI Tool drafts the supplier query or fraud alert, and an Evolus workflow routes the case. A clerk approves or rejects in a review queue, and every step is logged. The design keeps money safe. Code does every number. The agent can make a decision stricter but never looser, so it cannot approve a fraud finding. When its review doesn't come back, or a value can't be read, the invoice goes to a person with the reason shown. The model never sees the full bank account. Each finding shows its code, the rule that fired and the amount at risk, so a clerk can check it in seconds. The repo includes 208 offline tests, a Docker image built in CI, evaluation and benchmark scripts over 24 synthetic cases, and a demo mode that replays recorded runs, clearly labelled, if the services are down.
18 Oct 2026
