I build products with agents, and write down what breaks.
$ git diff --name-only $SHA..origin/mainapp/schemas/catalog.py$ ssh studio-warm-a 'git pull'$ alembic upgrade headA-HQCA, first author
Byzantine-robust federated learning and noise-aware QAOA for healthcare.
On the desk this month
Four things have my attention. The product pays for itself in lessons, the paper has a date on it, and the rest is writing.
- Cetus in production. Flow, Qwen and edge-tts behind one API.
- The A-HQCA paper, ready for Dubai on 25 November.
- This site, rebuilt from an empty repo.
- Five war stories written up, one per commit that cost me a week.
Five things I have shipped. Each one has a page with what it runs on and what broke.
cetus-one.vercel.app Cetus
One internal API in front of Flow, Qwen and edge-tts, so image, video and voice all run from a single composer.
- Writer10 scenes, routed through OmniRoute
- TTSedge_tts word boundary timestamps
- Alignerevery scene against 2.25 words per second
- EDLwritten from the JSON scene definitions
Kairo
A desktop app that writes a ten scene script, narrates it, checks the pacing, and hands back an edit list.
window 5 seconds, rollingfeatures 22+ per packetensemble random forest + xgboostreport rag over chroma, mapped to mitre att&ckfollow up 6 turns per incidentCyberBrain IDS
An ensemble reads live packets, a RAG pipeline writes the incident up against MITRE ATT&CK, and the analyst gets six follow up questions.
An organization adds its people.
An LLM writes the test.
It goes out from their own domain.
Answers come back graded.
Mailtrail
An organization adds its people, an LLM writes the test, it goes out from the organization's own domain, and the answers come back graded.
No checkout
The seller submits the device. The platform quotes a price. On acceptance, both sides carry on over email.
Console resale marketplace
A resale marketplace with no payment rail: the seller submits the device, the platform quotes, and on acceptance both sides carry on over email.
Adaptive Hybrid Quantum-Classical Framework with Byzantine-Robust Federated Learning and Noise-Aware QAOA for Trustworthy AI-Driven Healthcare
First author. 25 to 27 November 2026, to appear in AIP Conference Proceedings, indexed by Scopus and Web of Science.
- Operating room scheduling, optimality gap
- from to 0.9%
- Noise-Adaptive QAOA picking circuit depth per call from live IBM Eagle calibration. Runtime came down from 186s to 142s in the same run.
- Poisoning success rate
- from to 0.6%
- QDS and FLTrust as a two stage defence, at 2.1 ms overhead. 8 clients, 2 of them Byzantine, NIH ChestX-ray14 split non-IID.
- Genomic selection AUROC, TCGA-BRCA
- from to 0.901
- Against an L1 regularised logistic regression baseline, and on 30.4 selected features against the baseline's 47.3.
Write to me
Applied AI work, a product that needs building end to end, or a bug that has eaten a week of your time. I am in Lahore, and I have worked with clients across the USA and Europe.
shahzad.exec@gmail.com