Waseem
Hanif.

AI receptionist

An agent that actually books, not one that promises to

Most chatbot demos stop at conversation. The fan of tools on the right of that canvas is the difference: live calendar reads, appointment create, update and delete, a CRM lookup and append, and a confirmation it sends itself. Retrieval over the real service documents keeps its answers specific rather than generic.

The small chain on the left is the detail that makes it feel human. People send three short messages in a row, so it waits, flushes them together, and edits one reply into place instead of interrupting itself twice.

Tools

8

Channels

3

Sub-flows

7

n8n · Gemini · Pinecone · Google Calendar · Sheets CRM · WhatsApp · Telegram

GeoReport AI · conversational analyst

Describe it in a sentence, confirm before it runs

Say what you want mapped in plain language and it interviews you for whatever it still needs, proposes candidates when a place name is ambiguous, then shows exactly what it is about to run and waits.

That confirm step matters because every run costs real satellite processing time. Guessing at a vague request and mapping the wrong area is the expensive failure.

FastAPI · multi-turn agent · geocoding with candidate fallback · CesiumJS

Clipify · clip factory

Long video into short form, without losing the veto

A full web app, not a script. Paste a link and it reads the transcript first, so every moment is visible before a single byte of video downloads. You tick what is worth making, it renders with word-level captions, you watch each result, and only then does anything publish.

Reading the transcript first is also what makes it cheap. The download is only the chosen window, about 5 MB instead of gigabytes, and further clips from the same source cost nothing but a render.

Per build

~5 MB

Human gates

2

Cadence

4 h

Python · Flask · FFmpeg · Whisper · vision and transcript models · SQLite

GeoReport AI

Reports written from measurements, not guesses

Draw an area on a living globe, pick two periods, and get a change-detection report. The analysis runs first and the model only ever describes numbers it actually produced, so it can write the report but never invent the measurement.

Analyses

28

Turnaround

~10 min

FastAPI · Next.js · CesiumJS · Earth Engine · Sentinel · Landsat · MODIS

Text to published video

A scheduled pipeline reads a topic from a spreadsheet, has a model think through the treatment, generates the footage with Veo 3, uploads and publishes it, then writes the result back. Nobody touches it between topic and live video.

n8n · Gemini · Veo 3 Fast · Drive · Sheets

LeadGen Engine

Finds target businesses, reads each prospect’s site, writes a genuinely specific email rather than a template, and sends on a daily cap. Nothing leaves the machine until a person approves the draft. The engineering is in surviving unattended operation: dead connections reconnect rather than being misreported as bounces.

Python · Playwright · SQLite · SMTP and IMAP

Lumina

A conversational agent for night-sky and air-quality questions, holding six live data tools: VIIRS radiance, geocoding, weather, air quality and a NASA research index. It fetches real measurements rather than recalling them.

n8n · Gemini · VIIRS · OpenWeatherMap. The WebGIS platform it plugs into is a third-party open-source template; the agent is mine.

StickMind

Fully original animation with no third-party footage anywhere. The part worth showing is the validation harness: it re-derives the inverse kinematics and asserts feet-on-floor, no-slide and no-inverted-shin across every frame of every action, which caught four bugs that had shipped in every previous video.

Remotion · React · Whisper · inverse kinematics · FFmpeg

Trading systems

Multi-book MetaTrader 5 architecture, each book isolated on an always-on VM. The value is the apparatus rather than the strategy: walk-forward backtesting with separate in-sample and out-of-sample windows, half-Kelly sizing with a drawdown governor, circuit breakers. Every AI veto is forward-simulated into a shadow book so the layer can be measured instead of assumed. No performance claim is made here.

Python · MetaTrader 5 · PyTorch · LightGBM · Azure VM

Tooling

architect-agent interviews you about a project idea, then generates the architecture document, sub-agents and task breakdown. A Shopify scraper extracts structured product data and feeds the video pipeline above. A Meta Ad Library scraper pulls competitor creative, still in progress.

Python · Flask · Playwright · Docker · private repositories, walkthrough on request

Lunar south pole · 80 to 90°S

Artemis-class landing site screening

AHP-weighted fuzzy multi-criteria evaluation across NASA LOLA topography, illumination and roughness, with hard constraints vetoing unlandable terrain. The result was checked against an external benchmark rather than asserted: all nine NASA Artemis III candidate regions fall inside the top 0.1 percent of the scored surface.

Validation

9 / 9

Percentile

0.1 %

Grid

50 m

Python · rasterio · geopandas · scipy · NASA LOLA

ParcelAI · land record digitization

The honest limit that shaped the product

A multi-task U-Net trained on more than 20,000 real digitized parcels. Auto-accept tops out near 88 percent correct, so blind automation is not sellable and the project does not claim it. What is sellable is review order: worst-first review reaches 96 percent correctness at 70 percent of the effort, surfacing errors up to 2.5 times faster than random.

Parcel F1

0.689

Boundary F1

0.775

Tolerance

1.5 m

PyTorch · ONNX Runtime · MobileSAM · FastAPI · QGIS plugin