Founder & Lead Engineer at ProveMVP · Lahore, Pakistan
I turn startup ideas into working products. At ProveMVP we put a real product in front of real users in 14 days, then hand the founder a written keep, change, or stop. On the side, I go deep on how LLMs are served in production.
provemvp.com · info@provemvp.com · LinkedIn
2-Week MVP: one flow, a live URL, first users, and a decision by day 14. Code lives in the client's GitHub from the first commit. Pricing is on the site.
Recent work (case studies on provemvp.com):
| Product | What it is | Stack |
|---|---|---|
| Checkout Guard | Shopify app that blocks bad orders at checkout without code. Spec to production in 8 days, ~100 tests. | React Router, Shopify Functions (WASM), Prisma, Postgres |
| Family Handbook | Platform for families to write their values into a printed handbook, with chat, calling, and payments | Django REST, React, Stripe, Celery |
| Kharcha | Offline-first expense tracker in English and Urdu | Expo, React Native, SQLite, Node |
| Emma Vault | Website and subscription flow for an encrypted on-device vault | React, Stripe, Firebase |
| Elysium | Multi-store commerce backend and admin panel | Fastify, TypeScript, Postgres, Neo4j, Auth0 |
Open source from that work: checkout-rules-core, the typed rules engine behind Checkout Guard (Zod schemas, a compiler that fits rule sets under Shopify's 10 kB Function limit, an evaluator, 47 tests).
Stack I reach for: TypeScript, Next.js, React Native / Expo, NestJS, Node, .NET, Python / Django, PostgreSQL, Docker, AWS, Shopify.
A portfolio of hands-on projects on serving models, measuring them honestly, and making them faster and cheaper, all on free-tier T4 GPUs. Measured so far:
| Project | Headline result |
|---|---|
| servellm | Continuous batching on vLLM took aggregate throughput from 67 to 1,637 tok/s at 64 users (24x) while per-user ITL rose only 14.7 → 36.8 ms |
| tokenbench | Load-testing harness: found the saturation point where the server generated more tokens but goodput halved and only 29% of requests met SLO |
| kvcalc | Predicted KV cache capacity within 1.7% of vLLM's reported 340k tokens; prefix caching cut cache use from 24% to 5.4% |
| quantlab | INT8 and INT4 cut single-stream latency 2.4x and 3.2x; INT8 lossless, 4-bit costs 8–15 GSM8K points |
| specdec | Speculative decoding: k=1 gave +30% throughput at 4 users; k=5 was a net loss everywhere |
| fusedkernels | Hand-written Triton softmax beat torch.softmax by up to 35% (240 vs 178 GB/s) |
Every repo has a concept primer, CPU-runnable tests, the GPU notebooks that produced the numbers, and the runs that disproved the hypothesis.
- Have an idea to prove? ProveMVP builds it in 14 days. See pricing or book a 20-minute call.
- Need an engineer? I'm available for contract work and open to senior full-stack or LLM inference infrastructure roles, remote or in Lahore.