Jehanzeb Ali

From NFL on Roblox
to AI firewalls

Game developer and AI engineer in Karachi, shipping since 2019.

Get my CV

2017Unreal at Staffordshire

Six games for my BSc, solo and in teams of up to 19.

GVE showcase
3rd, collaborative games. 5th, game programming
Graduated
June 2020, BSc (Hons) Games Design and Programming

2020Four years of VR at Remio

Unity worlds for Meta Quest, for 100k+ monthly players.

Remio: team games and shared spaces.
Animal Blocks: three players fighting in a block forest
Animal Blocks, multiplayer sandbox.
Animal Blocks: a monster at night under gunfire
Survival nights.
On the Meta Store
15+ VR experiences
Loading times
Cut by 25%
Live ops
Monthly patches, Nov 2020 to Feb 2025

2025NFL Universe Football on Roblox

Lua and Luau at Voldex: gameplay, UI, live events and the store.

Six versus six, all 32 teams.

400K+

daily active players on the game I worked on

Loading screen art: a Ravens receiver making a one-handed catch
Mode loading screens.
Season 6: the new park.

2026TableSide, dishes in 3D

My company: phone photos of a dish become an AR menu in about ten minutes.

  1. Scan the plate

    Film the dish in the app. The pipeline builds the model, no manual modelling.

  2. See it on the table

    Diners scan a QR code and see the dish at true size.

  3. Order, pay, review

    Kiosk and table ordering, tied into the POS.

    130+dishes modelled
    5+restaurant brands
    4,000diner scans in 4 weeks

    tableside.pk

2026Valar, a firewall for AI

A self-hosted gateway between a company's apps and any LLM. Sensitive data is redacted or blocked before a prompt leaves the network.

Redact, block, audit. Built April to July 2026.

  1. One base URL

    OpenAI-compatible. Apps switch to Valar by changing one URL, with no SDK or code changes.

  2. Reversible redaction

    Values become typed placeholders like [REDACTED:pii.cnic:1] and are restored for the user afterwards. The model never sees them.

  3. Deterministic policies

    Versioned YAML policy packs decide observe, alert, redact or block. No second LLM, so every decision is explainable to compliance.

Valar risk dashboard: findings by type and the riskiest apps
Live risk across every app.
Valar audit record: request metadata, a blocked risk decision and masked snippets
Every decision explained, with no raw prompt kept.
Policy for secrets and CNIC numbers

Try a policy on a sample request, English and Roman Urdu

Sent by an employee

Is customer ka refund check karo. CNIC 42101-1234567-1, use key sk-example-9f2c on billing.example.com

2 findings: pii.cnic, secret.openai_key

Forwarded to the model

Is customer ka refund check karo. CNIC [REDACTED:pii.cnic:1], use key [REDACTED:secret.openai_key:1] on billing.example.com

Redacted. The reply comes back with the real values restored. The audit log keeps metadata only.

Detection
94% recall, 96% precision on a 400-prompt golden set
Latency
+7 ms at p95 before the model call
Detectors
20+ for PII, secrets and prompt injection
Security
Keys encrypted at rest, Argon2, metadata-only logs, every query scoped to an org
Tests
216, including proof raw secrets never reach the database
Stack
FastAPI, Postgres, Next.js, Docker, OpenRouter and LiteLLM
Lab
NVIDIA DGX Spark: LoRA and QLoRA fine-tuning, quantized local models behind Valar

valar.digital

Next, on the bench

Prototypes I am building toward launch.

  1. Ustaad

    O-Level Physics, twelve minutes a day.

    Mobile web app

  2. Lime

    Holdings and dividends for PSX investors.

    Mobile app

  3. Sudoku

    Puzzles graded by technique. No ads.

    iPhone and iPad

Games CV or AI CV

Same seven years, told for who is hiring.

Roblox, Unity VR, Unreal and mobile. Gameplay, performance and live ops.

Two pages, Roblox to Meta Quest.

Download games CV
Email
jehanzebali111@gmail.com
Phone
+92 321 2281328
Based in
Karachi, working remote
Download the reel