Amjad · Data science and AI engineering
Time series, NLP, and the plumbing that keeps a model breathing.
Two internships forecasting industrial SO2 emissions, tabular to sequential, then wired into predictive control.
On the weekend I make random software about some random stuff, or basically reverse engineer things nobody bothered to document.
Anyone can post a number and call it a win. I show what it beat. That's where the work begins. Every model looks like genius till you sit a dumb one down, and the dumb one ties the score. So who's wearing the crown?
class Amjad:
day = ["time series", "NLP", "the engineering that keeps it running"]
night = "taking apart hardware that shipped without software"
rule = "a result without its baseline means nothing"
def ship(self, model, baseline):
if model.score <= baseline.score:
raise ValueError("you learned autocorrelation, not the process")
return model, baseline # both. always both.graph LR
ME(("Amjad"))
ME --- ML["AI and Machine Learning"]
ML --- ML1["Time Series Forecasting"]
ML --- ML2["Model Evaluation"]
ML --- ML3["Computer Vision"]
ME --- NLP["NLP and LLM Agents"]
NLP --- N1["RAG and Vector Search"]
NLP --- N2["Multi Agent Systems"]
NLP --- N3["Text Classification"]
ME --- DE["Data Engineering"]
DE --- D1["Pipelines and Orchestration"]
DE --- D2["Docker and Object Storage"]
DE --- D3["SQL and Warehousing"]
ME --- RE["Reverse Engineering"]
RE --- R1["Bluetooth LE Protocols"]
RE --- R2["Packet Analysis"]
RE --- R3["Binary Inspection"]
classDef centre fill:#2C2624,stroke:#FF8A5B,stroke-width:3px,color:#F4F0EB
classDef ml fill:#3A2A20,stroke:#FF8A5B,stroke-width:2px,color:#FFD3BE
classDef nlp fill:#2A2438,stroke:#A98BE0,stroke-width:2px,color:#DCCEF5
classDef de fill:#1F2E2A,stroke:#86C08A,stroke-width:2px,color:#CFE8D1
classDef re fill:#33262B,stroke:#E5645E,stroke-width:2px,color:#F6CFCD
class ME centre
class ML,ML1,ML2,ML3 ml
class NLP,N1,N2,N3 nlp
class DE,D1,D2,D3 de
class RE,R1,R2,R3 re
| Project | What it does | Result |
|---|---|---|
| x20ctl reverse engineering |
My gamepad shipped with no PC software, so I decompiled the vendor app, recovered the Bluetooth protocol, and wrote the tool that should have existed | Protocol documented from scratch. No prior public record of it exists |
| sma-veille-technologique LLM agents |
Three agents read AI publications overnight and write the summary | Daily report in under 2s, no human in the loop |
| spam-detection-nlp NLP |
Sorts email into legitimate and spam, three approaches put head to head | 99% correct on 1,160 test messages |
| weather-data-pipeline data engineering |
Medallion pipeline over 10 cities, bronze through gold | Automated end to end, ingestion through dashboard |
| pneumonia-detection-pytorch computer vision |
Spots pneumonia in chest X-rays | 87.5% on 624 images, against a 62.5% baseline |
| arima-vs-lstm-forecasting forecasting |
Classical statistics against deep learning, same data, same split | Compared on 43,848 hourly readings |
Machine learning
Data engineering and infrastructure
Also in the bag
A baseline is a bar. Something already standing in the room, doing the job badly, waiting to see whether you can clear it.
This field makes it easy to fool yourself. A model can look brilliant while it is only repeating the last value it saw. An accuracy can look strong until you notice one class is 90% of the data. A metric can climb while the thing you actually cared about goes nowhere. The only defence is to stand something stupid next to your work and check that you genuinely beat it.
So every repo here carries what it was measured against, and says where the method breaks down. Not out of modesty. A number with nothing beside it isn't evidence, it's just a claim.
Anyone can post a number. Show me what it beat.
Open an issue on any repo if you want to talk.