Lamis Technologies

Projects / demos

Projects

Personal projects I build to try out ideas end to end, from data to a small working interface. All the code is on GitHub.

01 / Similarity search · geospatial

HikesAlike

Pick a hike you enjoyed and HikesAlike lists other hikes in Europe that feel similar: comparable length, climb, highest point, landscape and typical weather. You can also describe a hike instead of naming one, and limit results to a driving time from a starting town.

  • Builds a catalogue of roughly 250,000 day hikes and day stages from OpenStreetMap, enriched with elevation, land cover and climate data.
  • Each hike becomes a feature vector; similarity weights are adjustable in the interface.
  • Optional drive-time filter through a self-hosted Valhalla routing engine.
  • Served as a small read-only export on Google Cloud Run.

Technologies

  • Python
  • OpenStreetMap
  • SQLite
  • Valhalla
  • Leaflet
  • Cloud Run
Open live demo Project page Code on GitHub

The server sleeps when idle, so the first load can take around 20 seconds.

HikesAlike page for the Aljažev dom to Triglav hike, showing the route on a map, its elevation profile and summary chips for distance and climb.
fig. 1.1 · HikesAlike
Table of ten hikes similar to the climb to Triglav, each with a similarity score and coloured cells for how alike it is on length and climbing.
fig. 1.2 · HikesAlike
02 / Simulation · inventory policy · RL

Supply chain simulator

A small multi-echelon retail supply chain simulator. Factories, warehouses, shops and customers form a directed graph that shares one stochastic world with seasonality and disruptions. Each node runs its own replenishment policy, so ordering rules can be compared on identical random worlds.

  • Textbook inventory policies — (s,S), (s,Q), (R,S) and (R,s,S) — extended to route orders across several suppliers.
  • A PPO reinforcement learning policy trained against the textbook baseline with common random numbers, plus an Optuna tuner.
  • Optional LLM world generator and replay of real demand from the M5 (Walmart) dataset.
  • Work in progress on JEPA-style world-model policies on experimental branches.
  • A limited Streamlit demo compares four policies on one shop and includes a 30-world check of whether a gap is luck.

Technologies

  • Python
  • PyTorch
  • PPO
  • Optuna
  • Streamlit
  • Cloud Run
Open live demo Project page Code on GitHub

A limited slice of the simulator. It may take a few seconds to wake up.

Streamlit supply-chain policy lab showing a scorecard that compares four inventory policies on net profit, fill rate and days out of stock.
fig. 2.1 · Supply chain simulator
Bar charts comparing a reinforcement learning policy with an order-up-to baseline across 32 paired seeds on service level, stockout rate, turnover, price and revenue.
fig. 2.2 · Supply chain simulator
03 / Computational geometry

KrovLab

KrovLab takes a building footprint and a roof pitch and returns the geometry of a hipped roof: faces, hips, ridges, valleys and the quantities a roofer needs. When a plan cannot be roofed it returns a named failure rather than a wrong answer.

  • Weighted straight-skeleton core with no third-party dependencies.
  • Supports gables, knee walls, gambrels, courtyards, overhangs, several cells per project and dormers.
  • A Flask web demo draws the roof plan and a 3D view, accepts DXF footprints and exports OBJ and GLB meshes.
  • Its limitations are documented up front.

Technologies

  • Python
  • Computational geometry
  • Plotly
  • Flask
  • DXF
Open live demo Project page Code on GitHub

The server sleeps when idle, so the first load can take a few seconds.

Roof plan of an L-shaped building at 45 degrees, with hips in orange, ridges in red, a valley in blue and heights at each vertex.
fig. 3.1 · KrovLab
Three-dimensional view of the same L-shaped hipped roof with its ridge and hip lines.
fig. 3.2 · KrovLab
04 / Optimisation · energy · LangGraph

AeroGrid

A research demo of a home energy planner. A streaming agent shifts deferrable household loads (EV charger, hot-water heater, dishwasher, washing machine) and dispatches a home battery into cheap price slots, using real SMARD wholesale prices for Germany and Luxembourg.

  • Receding-horizon optimiser: a linear programme, or a small mixed-integer programme when an appliance cycle is waiting for a decision, solved with HiGHS through CVXPY.
  • LangGraph loop that replans on new appliance starts, price surprises, EV deadline slips and a 15-minute resync, with a human approval step before a plan is committed.
  • A 1 Hz digital twin runs a no-optimisation baseline and optimiser variants with and without the battery on the same prices.
  • Known gaps are listed in the README: rectangular appliance cycles, a naive price forecast and no solar or grid export.

Technologies

  • Python
  • LangGraph
  • CVXPY
  • HiGHS
  • SMARD
  • Jupyter
Open live demo Code on GitHub

The demo is a replay of recorded simulation runs. The planner itself runs locally from the command line and notebooks.

Diagram of the AeroGrid LangGraph nodes: forecast price, optimise, propose reschedule, human approval gate and commit plan.
fig. 4.1 · AeroGrid
Charts of one optimiser scenario with an 8 kW house power cap, where the EV charger and the water heater share the cheapest slots to meet the EV deadline.
fig. 4.2 · AeroGrid
05 / LLM agents · LangGraph

CV writer

An agentic workflow that drafts a CV and cover letter for a specific job. It reads your existing documents, extracts the job requirements, researches the company, retrieves relevant past projects and iterates on drafts with a critique step, pausing for your input at each decision.

  • LangGraph state machine with separate agents for reading, research, retrieval, routing, writing and critique.
  • Two-stage retrieval: vector similarity followed by an LLM relevance check.
  • Drafts are scored and refined until they pass a quality threshold; every version is kept.
  • PDFs are produced through LLM-written LaTeX, with compiler errors fed back for repair.

Technologies

  • Python
  • LangGraph
  • LangChain
  • RAG
  • Tavily
  • LaTeX
Project page Code on GitHub

Runs locally from the command line; there is no hosted demo.

LangGraph workflow diagram: load documents, analyse the job description, research the company and retrieve experience, then a router that loops through drafting, critique and PDF generation for the CV and the cover letter.
fig. 5.1 · CV writer