Software Engineer

Arian Boukani

I build real-time backends and AI pipelines.

  1. 01
    SevenTask
    Real-time task app for teams
  2. 02
    OptiServe
    Tunes serverless ML pipelines
  3. 03
    Landline
    Voice agent for apartment hunting
  4. 04
    Guardian
    Health agents on a local model
  5. 05
    BakeryPilot
    AI assistant for bakery operations
MSc, Systems & Performance Engineering 2arian3.github.io github.com/2arian3

01Technical co-founder · 18 months live

SevenTask

A real-time task and project app for teams.

  • Boards update live when a teammate changes something
  • Time tracking on every task
  • Chat and calendar are built in
  • Each task keeps its full edit history
200+
monthly active users
AngularASP.NET CoreSignalRSQL ServerEF CorePrometheus
A board with one team's own columns.
The board view, from the product.
SevenTask · System design

How edits reach every open board

flowchart TB
  subgraph CL["Clients"]
    direction LR
    W["Angular SPA · SSR"]
    A["Admin app"]
  end
  subgraph API["ASP.NET Core"]
    direction LR
    R["REST · field-level writes"]
    H["SignalR hub · typed events"]
  end
  subgraph DATA["Data"]
    direction LR
    D["Domain package · EF Core"] --> DB[("SQL Server")]
  end
  CL --> API
  API --> DATA
  • Fields save separately, so two people can edit one task.
  • SignalR pushes each change to every open tab.
  • Cards move right away, and move back if saving fails.
SevenTask · Workflow engine

Each team sets up its own columns

flowchart TB
  subgraph G["Guards, in order"]
    direction LR
    G1["Task exists"] --> G2["Caller has access"] --> G3["Column is on this board"]
  end
  subgraph W["One transaction"]
    direction LR
    W1["Update the task"] --> W2["Record the move"] --> W3["Log the event"]
  end
  G --> W
  • Pilot teams all wanted different columns, so they're stored as data.
  • Each project or team can have its own set.
  • The server rejects a move into another board's column.
  • Every move is logged with who made it and when.
SevenTask · Demo

A tour of the live product

Demo Tasks, teams, chat, calendar docs/seventask-demo.gif
  1. My tasks, as a list and as a board.
  2. Team pages with departments and project dependencies.
  3. Direct messages and the shared calendar.

02MSc thesis · solo

OptiServe

Chooses memory size and model for serverless ML pipelines.

  • Runs your Lambda functions at different memory sizes
  • Handles pipelines with branches and retries
  • You pick the goal: speed, cost or accuracy
103,680
possible setups; the search tries under 200 of them
PythonAWS LambdaStep FunctionsCloudWatchSciPymoto
Sweet spot Memory → Cost per run
Cost per run falls, then rises, with memory.
The chosen setup, checked against brute force.
OptiServe · System design

How OptiServe picks a setup

flowchart LR
  subgraph ON["Online · real AWS"]
    direction TB
    L["Live Lambda"] --> P["Profiler"] --> F["Curve fit"]
  end
  subgraph OFF["Offline · pure math"]
    direction TB
    W["Workflow model"] --> O["Greedy optimizer"] --> R["Best config"]
  end
  ON -->|"lookup table"| OFF
  • It runs the real function at several memory sizes.
  • Fitted curves predict latency and cost for the whole pipeline.
  • A greedy search then follows the critical path.
  • Only the first step calls AWS. The rest runs offline.
OptiServe · Demo

The search, scored against brute force

Demo Pipeline, trade-off, search, brute-force check docs/optiserve-demo.gif
  1. The pipeline: six Lambda steps with branches and retries.
  2. Set a budget and a quality floor.
  3. Watch the search pick a setup, then the brute-force score.

03Boson AI hackathon · data, ranking, orchestration

Landline

A voice agent that phones listing agents for renters.

  • Call it in any language and say what you want
  • A live map of Toronto listings re-ranks as you talk
  • After you hang up, it phones every place you picked
  • Every fact on the map shows who said it
3rd
place out of 30 teams
Higgs RealtimePythonFastAPIPostGISTwilioNext.js
A listing card after the agent called.
Call results landing on the live map.
Landline · System design

Two flows, joined by one frozen call plan

flowchart TB
  subgraph F1["Flow 1 · the renter's call"]
    direction LR
    R["Renter · phone or browser"] --> CO["Concierge · Higgs Realtime"]
  end
  subgraph MINE["Plain code · my part"]
    direction LR
    O["Orchestrator · one writer per session"] --> M["Matcher · ranks listings"]
    M --> DB[("Postgres + PostGIS")]
  end
  subgraph F2["Flow 2 · the listing calls"]
    direction LR
    D["Dispatcher"] --> CA["Callers x N · Higgs Realtime"] --> AU["Auditor · quoted facts only"]
  end
  F1 -->|"tool calls"| MINE
  MINE -->|"frozen call plan"| F2
  • Models handle speech. Code ranks, dials and books.
  • The renter's call ends with one frozen call plan.
  • Ranking 4,000 listings takes about 2 milliseconds.
  • A fact shows only with the agent's exact words.
Landline · Demo

The live map while the calls run

Demo Map re-ranks, calls go out, results land docs/landline-demo.mp4
  1. A lower budget re-ranks the numbered pins.
  2. After the hang-up, the picked places are called at once.
  3. Results land: a hidden fee, a scam flag, a booking.

04NVIDIA Spark Hack Toronto · model layer

Guardian

Moved a six-agent health app onto one desktop GPU.

  • A watch tracks heart rate and detects falls
  • Agents talk to the patient through the watch
  • No patient data goes to the cloud
30B
parameters, about 3B active per token
TensorRT-LLMNVFP4Nemotron NanoFastAPILangGraphDocker
Each token uses only a few experts.
The dashboard as it calls 911.
Guardian · System design

All model calls go through one interface

flowchart TB
  C["Watch · phone · web"] --> APP
  subgraph APP["App · built by the team"]
    direction LR
    A["FastAPI + router"] --> G["6 agents"]
  end
  subgraph MINE["Model layer · my part"]
    direction LR
    S["LLM interface"] --> T["TensorRT-LLM · NVFP4"]
    S -.-> O["Cloud API"]
  end
  APP --> MINE
  • No agent imports a vendor's library.
  • 4-bit weights: ~16 GB instead of ~60 GB.
  • Moving from cloud to local was a config change.
  • The memory budget leaves room for a second model.
Guardian · Demo

A fall on the watch becomes a 911 call

Demo Vitals, a fall, a check-in, a 911 call docs/guardian-demo.gif
  1. The watch streams heart rate, and risk rises.
  2. After a fall it asks "Are you okay?"
  3. No answer, so it calls 911 and the family.

05TMLS Hackathon · data, proxy, deploys

BakeryPilot

An AI assistant for a bakery supply chain.

  • Ask about stock and orders in plain English
  • Agent changes wait for someone to approve them
FastAPILangGraphPostgreSQLNginxNext.jsGitHub Actions
An agent's change, waiting for approval.
Orders flowing across the supply map.
BakeryPilot · System design

The agent uses the same API as people

flowchart TB
  subgraph REQ["Ask"]
    direction LR
    U["Operator"] --> N["Nginx"] --> B["FastAPI"]
    B <--> A["LangGraph agents"]
  end
  subgraph GATE["Approve"]
    direction LR
    C["Action card · pending"] --> H["Person approves"] --> D[("PostgreSQL")]
  end
  REQ --> GATE
  • Agents get no direct database access.
  • A change the agent wants waits as a pending card.
  • Approving saves the change and its audit record together.
  • Postgres triggers stop anyone editing history.
BakeryPilot · Demo

The supply map and the copilot

Demo Supply map, copilot, order flows docs/bakerypilot-demo.gif
  1. The supply map flags a supplier at risk.
  2. One copilot request creates sample orders and schedules.
  3. Expiring lots show up as alerts.

Questions?

Happy to go deeper on any project.

Arian Boukani 2arian3.github.io github.com/2arian3 Deck: 2arian3.github.io/projects