Ahmed Elmalt
Principal Engineer·
I am a Principal Engineer at Amazon, currently the lead PE for Alexa's communications organisation: three pillars, 17 teams, 150+ engineers, 107 cloud services. Most of my time goes on real-time multimodal systems, agentic architecture, and getting AI to behave predictably at scale. This blog is where I write about the engineering behind it.
How I got here
I have been building production systems since 2003, and the thread running through all of it is platforms: the layer other engineers build on. I started in Cairo, on the interpreter that turned OmniPage's OCR output into Open XML, and an e-government portal serving 30M+ citizens. At GE Healthcare I built the location infrastructure for hospitals, fusing radio, infrared and ultrasound until it was accurate to under 2%, which took equipment search from 24 minutes down to 4. At Cornerstone OnDemand I designed a platform language, wrote its compiler, then built the virtual machine that ran it. At GoFormz I was employee number nine and grew the engineering team to 25+ through two funding rounds.
Eight years of it at Amazon, all on Alexa. Two habits survived the whole way. I settle the contract first, because an abstraction chosen badly is the expensive kind of mistake and you pay for it for years. Then I build the risky part myself before a team commits to it. That is how I know a managed language can sit beside a real-time media path, and that the dominant cost in visual AI is invoking a model per frame rather than the reasoning everyone assumes.
Stuff I built recently
Real-time visual intelligence
Alexa's platform for understanding what a camera sees, on smart eyewear and on screens. I set two abstractions and then defended them: one normalised output no matter which processor produced it, and processors that plug in anywhere from the device to the cloud. I wrote the first version of the memory service myself. Retrieval was the part most likely to fail, and I wanted to find that out before a team was staffed against it.
15% → 36%multi-hour recall
The real-time media platform
Every real-time call on Alexa runs through this, and it used to take a media specialist weeks to build anything on it. I rewrote it end to end. A working application is eight lines now because of one decision: the engine stays native and each SDK is a thin wrapper over it, so the media path never waits on application code.
1,100 → 8lines, and weeks to an hour
Calling from a billion-user platform, on Alexa devices
Two companies' code in one process, on hardware with a memory budget measured in megabytes. The plan I inherited was glue written once per operating system. I argued for a single interface contract compiled to each instead, then cross-compiled the partner's toolchain onto the smallest device we shipped, because a build path either exists or it does not and nobody had checked.
40+ devicesthree operating systems, one contract
Agentic incoming communications
A supervisor agent plans the work; domain agents only ever speak through an event bus. That looks like ceremony until the second channel arrives and nothing has to be rewritten. I built the prototypes first, including the synthetic mail pipeline, since none of this could be developed against real customer messages.
<2s · >85%latency, accuracy
Alexa's first contextual targeting
Semantic reasoning over a knowledge graph for the relationships a model cannot see, BERT scoring for the language it can, and an inference path that had to answer inside a hundred milliseconds or not at all. It became the foundation the advertising business was built on.
100+ TPS75ms p50, 99.9% up
A dialog policy that beat the rules
A reinforcement learning agent for skill discovery, up against a hand-written policy that was harder to beat than it sounds. It won on success rate and on dialog length at the same time, which I did not expect. Published, so the claim is checkable rather than mine to make.
73.4% → 77.0%arXiv:1911.06747
Also
- API bar raiser for Alexa's public camera, contacts and communications specifications.
- Patent filed, first-named inventor of five.
- Twelve engineers mentored through formal programmes, several now architecture owners.
- Host of Kalam Falsfa, the Middle East's leading philosophy podcast.