I'm Rehan, a full-stack AI engineer in Karachi and an MS AI student. I ship RAG and document-AI pipelines, LLM apps and the product around them: 100+ freelance orders delivered, on-device AI in apps people use every day, open-source infrastructure, and a co-authored paper under review.
100% five-star across 69 client reviews from 19 countries
Fig. 01 · Rehan, Karachi24.86° N, 67.01° E
RAGNext.jsFlutterPythonCurrently
MS Artificial Intelligence
Bahria University · 3rd semester
01Selected work
Client systems, AI products, mobile apps and open source. Every number on this page links back to where it came from.
My mother tongue had no search engine for its greatest poet, and Roman Sindhi has no standard spelling. Could retrieval still find the right verse?
A retrieval-augmented search engine over Shah Jo Risalo, the 18th-century Sindhi Sufi poetry of Shah Abdul Latif Bhittai.
2,519
verse documents indexed
FAISS IndexFlatL2, 384-d multilingual embeddings
Buyers and sellers of physical commodities needed live prices, matched orders and payments in one place, not spreadsheets and email.
A real-time global commodity trading platform with a matching engine, carbon-credit listings, live WebSocket data and automated trading bots.
38+
REST endpoints
Plus 8 auth routes and a dedicated WebSocket channel
Rate limiting should not cost more latency than the request it protects.
A framework-agnostic rate-limiting library for Node.js and Bun: 8 framework adapters, 8 storage backends, millions of checks per second.
5.5M+
ops/sec on Bun
4M+ on Node.js, 10K-key benchmark
Searching your own gallery shouldn’t mean uploading it to someone else’s server.
An iOS-style photo gallery for Android that understands your photos with on-device AI, without sending anything anywhere.
0
network permissions
No Internet access, accounts, ads or analytics
A Maps list tells you a business exists. It doesn’t tell you what it’s struggling with, or who to call about it.
Type “Cafes in Dayton, Ohio” and get sales intelligence back: every business found, enriched, audited and prioritized, with a client-ready audit PDF for each.
8
pipeline stages
Discover to export, checkpointed item by item
Most scrapers are one-off scripts. Making one safe to run for strangers, on any URL they type, is a different problem.
A generic web scraper SaaS: point it at a URL, it detects listings and pagination, you map fields, and it exports clean data. Hardened for multi-tenant use.
~4×
faster detail-page fetching
Bounded concurrency instead of a sequential loop
Also on the shelf
02Clients, verbatim
Pulled straight from Fiverr: 69 reviews from 19 countries. Typos included. On Fiverr I’m rehan_sathio.
i had another guy working on the same project for almost a month and failed! rehan did it in three days!!!! … Anything he didn't know, he quickly researched and found a resolution immediately.
vasilynaumenkoUnited States · Fiverr · 3 years agoHe immediately understands requirements, and is very tech savvy. … he is not afraid of doing more than asked
hit him up late at night, answered right away. Told him the requirements and he started working that night. By the morning, he's done more than 40 % of the work.
This is the fifth time I’ve worked with this developer, and I’m still amazed by his expertise. The website’s AI and data science features are exactly what I needed, and the execution was perfect.
paulstank347Finland · Fiverr · 10 months agoFifth project togetherWas having a small melt down, was able to get the fix i need to continue on with my own project.. very patient and would recommend.
rehan stayed up all night and finished the whole react dashboard. Very impressive work! I'd definitely work with rehan again.
Asked to clone and teach me a complex website, with the algorithm! insane.
metoldmeGuam · Fiverr · 3 years agoWhere the reviews come from
Reviews by country
amazing developer !!! made sure everything was done exactly how i needed it and was quick and efficient with communication and delivery times !!! i had given him more requirements and revisions multiple times, not once did he complain or get mad, just simply made sure that i was accurate with the requirement so he can make it quickly !!! 10000% recommend him for anyone and everyone !! literally couldnt ask for anyone better in anyway will 100% be coming back to him again !!! thank you very much !!!!
sxa_07FiverrGBI very thankful with Rehan. He delivered all as expected and right on time. Communication is very good and the output he delivered has no problem at all. I highly recommended his service and i am pretty sure to buy from him again. Thank you and god bless.
Excellent work. Really makes an effort to deliver outstanding websites that comply to all your wishes. Rehan is the peak of customer service. 10/10 experience and would come back again in a heartbeat.
arnaiseFiverrBEThe website satisfied all of my requirements, and any changes i requested were implemented promptly. Along with the construction, i was given a lot of explanations of various ideas to make the process go smoothly. Very pleased and will place another order for next projects.
Consistently impressed with the quality of work! Every time I work with Rehan, the project is delivered on time and with exceptional attention to detail. The AI integration is always flawless.
paulstank347FiverrFIrehan did a very good job and the result not only met my requirements but even exceeded them. he thinks along and gives helpful suggestions. Communication with him was pleasant and very professional. I will definitely work with him again. Highly recommended.
03The proof
If a number can’t be checked, it isn’t here. Private client work is marked as private, and I’m happy to walk you through it.
80+
five-star reviews
From 100+ completed Fiverr orders. Level 2 seller. One client came back for a fifth project.
27.45%
Recall@10, measured
On my own 51-query benchmark for Sindhi poetry retrieval. Modest, real and reproducible.
38+
API endpoints in production
Commodity exchange: matching engine, WebSockets, Stripe + Binance/TRC20, CI/CD to a VPS.
118
students on my LMS app
Across 5 Bahria University campuses, with no institutional backing. 24.58% monthly active, from my own analytics.
52.31%
graph edges pruned
Co-authored trajectory-prediction paper, submitted and under review. minADE5 0.3863 m on nuScenes.
97
papers mapped
Three literature syntheses (31, 33 and 33 works) across document AI, few-shot AR and agent reliability.
8/8
invoices extracted
Hybrid OCR + LLM pipeline, 100% success on its production pilot.
5.5M+
rate-limit checks / sec
Hitlimit on Bun (4M+ on Node.js), 10K-key benchmark. My open-source rate limiter.
Ignite
grant for ChatDoc
My final-year project, an NLP + face recognition doctor's assistant, was funded and showcased.
3.65
MS AI CGPA
Bahria University Karachi, as of Spring 2026. 3.62 in my BSCS, with the President’s Honor Roll in Spring 2020.
75%
merit scholarship
BSCS at Mohammad Ali Jinnah University, 2019–2023.
Lead
GDSC MAJU Web Dev, 2022–23
Google Developer Student Club. Seminars, competitions and freelancing awareness sessions.
04Capabilities
Engineers who can’t do research, researchers who can’t ship. I’m trying hard to be neither.
01
Retrieval, extraction and LLM systems that hold up outside a notebook.
Proof → Ask Bhittai · Invoice Intelligence · AIPhotos
02
From the database schema to the last pixel, on web and phone, on stacks clients actually run.
Proof → 100+ Fiverr orders · Commodity exchange · 14+ Flutter apps
03
Asking whether it works and proving it with numbers.
Proof → 1 paper under review · 97 papers mapped · 51-query benchmark
04
I hate repetition, so I automate it: workflows, deploys, generators, bots and the boring parts.
Proof → Invoice approval workflow · trading bots on cron · own dev tools
05Research
A co-authored paper under review, benchmarks, experiments and literature work from my MS in Artificial Intelligence. Every claim comes with a number or a source.
62%caught, where 90% was the target
MS Artificial Intelligence thesis, Bahria University
Teachable recognizers for blind and low-vision users always name an object, even when it isn’t there. In a pilot on ORBIT-India, thresholds meant to catch 90% of appearances caught 62% at best. The thesis calibrates that decision.
27.45%Recall@10
A 51-query retrieval benchmark over Shah Jo Risalo. Recall@1 9.80%, Recall@5 23.53%, Recall@10 27.45%: honest baselines on a hard, low-resource corpus.
92.86%precision
MS AI project, Bahria University
XGBoost + SMOTE for the statistics, an OWL 2 ontology for the reasoning. 92.86% precision and a plain-language reason for every risk decision.
06Status board
Third semester of the MS, laptop open, shipping in public.
Last updated
Researching
Phones that help blind and low-vision users find their own objects always name something, even in an empty frame. My pilot on ORBIT-India: thresholds meant to catch 90% of appearances caught 62% at best. I’m building a conformal calibration method that tells the user “it’s not here” at a tested level of reliability, on a mid-range Android.
Learning
My favourite course so far. The part that clicks is the math: how weights actually move and why a model works, not just which API to call.
Building
An iOS-style Android gallery with EmbeddingGemma running fully on-device: natural-language search across photos, recordings and video moments. I use it daily; the Play Store is next.
Learning
Preparing for the MLA-C02 exam, to back the deployment side of my AI work with a real credential.
Shipping
Short reels around the projects I build, posted while they happen.
Available
Full-time AI / full-stack / backend roles, remote or on-site in Karachi. Freelance projects too.
07Live
GitHub contributions, year by year, and everything that’s live right now. Open any of it and try it.
346 contributions in 2026
08 · Contact