The AI laboratory
The AI laboratory
The AI laboratory
Four of these modules compute real results in your browser — k-means, Holt–Winters, BM25 retrieval and ordinary least squares — on data whose origin is stated on each card. Two are architecture visualisations, and they say so. Nothing here fabricates an output or claims a connection it does not have.
Module 01
k-means++ over a two-dimensional customer space: recency against monetary value. The same mechanic produces the RFM segments inside Fidras, where the axes are engineered features rather than two toy dimensions.
●Real k-means++ running in your browser on generated example data. Centroids, inertia and segment sizes are computed, not scripted.
Inertia
43775
Within-cluster sum of squares
Iterations
4
Until assignments stabilised
Segment sizes
Raising k always lowers inertia — which is exactly why inertia alone cannot choose it. In production the segment count is decided by whether a human can describe each group in one sentence and act on it differently.
Module 02
Additive Holt–Winters with weekly seasonality, intervals from in-sample residual spread, and the inventory decision derived from the distribution: safety stock, reorder point and stockout probability.
●Real model fitted in your browser to clearly-labelled example SKU history. The production TCG system replaces this baseline with LightGBM over engineered features — the decision layer on top is the same maths.
Demand over lead time
1333
7 days ahead
Safety stock
129
z = 1.64
Reorder point
1462
Order when stock crosses this
Stockout risk
19.6%
With 1400 on hand
In-sample MAPE
15.5%
Fitted vs actual
Residual σ
29.7
Drives interval width
Push the service level to 99% and watch safety stock climb while stockout risk falls. That trade is the actual decision — a single point forecast cannot express it, which is why nothing I ship returns one.
Module 03
Ask a question about my work. The explorer tokenises the query, scores every chunk of this site with BM25, shows you exactly what was retrieved and why, then answers using only sentences from those sources.
●Real lexical retrieval over this site's own content, computed in your browser. The answer is extractive — no model call, no generation, so nothing can be hallucinated. Production systems use embeddings in pgvector plus a grounded LLM.
Retrieved context
Profile
score 7.64
Syed Nassirodin Quadri is an AI Engineer and Data Scientist based in Dubai, U.A.E. AI Engineer with an MSc in Data Science and production experience shipping Generative AI, Machine Learning and Data Analytics systems across SaaS, ecommerce, marketing technology, voice automation and enterprise operations. He holds a UAE Golden Visa — valid to 2036.
VesselReach — Contribution
score 6.72
On VesselReach, Syed's own contribution: Designed the whole system: ingestion, enrichment, scoring, outreach and the human handoff boundary. Built the data model and the enrichment pipeline, including verification and confidence scoring. Implemented the agent layer and the guardrails that stop it contacting the same company twice or acting on low-confidence data. Shipped and operate the production deployment.
ATS Resume Generator — Problem
score 5.68
ATS Resume Generator: Applicant tracking systems reject well-qualified candidates on formatting and keyword mismatch. Tailoring a resume per role is the right answer and nobody has time to do it ten times a week.
Stack — AI & LLMs
score 5.40
AI & LLMs: GPT, Claude, Gemini, DeepSeek, LLaMA, Mistral, LangChain, RAG, Hugging Face, pgvector.
Extractive answer
Shipped and operate the production deployment.
— VesselReach — Contribution
AI Engineer with an MSc in Data Science and production experience shipping Generative AI, Machine Learning and Data Analytics systems across SaaS, ecommerce, marketing technology, voice automation and enterprise operations.
— Profile
Syed Nassirodin Quadri is an AI Engineer and Data Scientist based in Dubai, U.A.E.
— Profile
Corpus: 60 chunks drawn from the project pages, case studies, architectures, experience and insights on this site. Every sentence above is quoted from one of them, with its source named — the same citation discipline I build into production RAG.
Module 04
Step through the outreach agent's state machine: what it does at each state, which guard has to pass, and the point where it hands control to a human.
●Architecture visualisation of the production VesselReach agent. It sends nothing, contacts nobody and calls no model from this page. The guards described are the guards that run in production.
Ingest port-call signal
guard: Signal resolves to a known vessel and window
An arrival event arrives from the collectors and is normalised into a call record.
Module 05
Tap the plot to add an observation. The line is refitted by ordinary least squares on every change, with R² recomputed. Add an outlier in the corner and watch how little it takes to move a fit.
●Real OLS computed in your browser from the points you place. No pretrained model, no stored result.
Slope
0.662
Intercept
16.69
R²
0.905
Variance explained
Observations
40
R² rewards a line for following the data it was fitted on, which is why I never ship it as evidence on its own. In the forecasting work the equivalent check is time-based validation: fit on the past, score on a period the model never saw.
Module 06
A realtime multilingual voice agent for inbound and outbound calls: qualification, grounded FAQ answering, routing, booking, transcription and WhatsApp follow-up.
●No live call backend is connected to this website, so there is no demo to press. Telephony, streaming speech and model credentials belong on a server, never in a browser. What follows is the production architecture.
Status
Demo offline on this site. If you would like to hear the agent handle a call, ask me for a scheduled walkthrough and I will run it against a live deployment.
Request a walkthroughProduction architecture
Ingestion
Processing
Models
Orchestration
Storage
Interfaces
Select any component to inspect it. Architecture of the production system. Nothing on this website places or receives calls.