Hire Ibrahim Salman
You work directly with Muhammad Ibrahim Salman (Ibrahim Salman), the engineer who builds it. I make grounded RAG assistants, OCR for scanned and multi-column pages, monitors that fail closed, and the Next.js and Python apps around them.
- Based in
- Islamabad, Pakistan, UTC+5. Working remotely.
- Engagements
- Fixed scope, sprints or a one-off audit.
- Payment
- Fixed scope: half at kickoff, half at handover.
- Proof
- 914 collected tests, public on GitHub.
What you can hire me to build
AI assistants that answer from your documents
Retrieval-augmented generation (RAG) over your documentation or website: crawled with Crawl4AI, chunked, searched with Convex vector search and reranked, then answered through the Vercel AI SDK with a fallback model (gpt-oss-20b on Groq, then Gemini). When your retrieval is empty, the routing selects refusal; low scores select hedging. These instructions still need evaluation against your questions.
UET GPT, live and unofficialCase studyService overview
OCR for scanned and multi-column pages
Text and tables out of scanned books, multi-column PDFs and slides, with RapidOCR on ONNX and PyMuPDF. I score OCR output against hand transcriptions: B.L.A.S.T.'s character error rate went from 0.499 to 0.108 on 14 hand-checked pages from one 1959 book, and the published comparison shows where Tesseract still does better.
B.L.A.S.T., 914 collected testsCase studyService overview
A feed that won't load looks exactly like “no new jobs”.
Monitors and alerts that fail closed
Scrapers and ETL jobs in Python with HTTP/2 workers, Pydantic validation and SQLite in WAL mode. A source that stops loading is recorded as degraded, with the reason, never as empty. The marketplace monitor I built during my internship at Paramount Intelligence shipped with 304 collected tests across a CI matrix of Python 3.10 to 3.13 on Ubuntu and Windows.
Marketplace Monitor, 304 collected testsCase studyService overview
Web apps in Next.js and Python
Next.js App Router with TypeScript in strict mode and Tailwind, Convex or a Python service behind it, and tests that run in GitHub Actions. Most projects are one of the three systems above; the web app is how people use it.
UET GPT and this portfolioView all projects
Tell me what you're building
A few lines is enough. It goes straight to me, and I reply by email. Or write to ibrahim.pk848@gmail.com.
Get a rough scope
Pick what you're building and how far it should go. You'll see the pipeline, the stack and a typical timeline. The real quote comes after a call.
2-4 weeks
Typical timeline for this build
An assistant over your docs and sites, with citations, refusal instructions for empty retrieval and a fallback model.
How it fits together
- Docs/Crawl
- Chunking
- Vector Search
- Cross-Encoder Rerank
- Multi-LLM Router
- Streaming UI
Stack
- Next.js 15
- TypeScript
- Convex Vector Search
- Vercel AI SDK
- gpt-oss-20b on Groq / Gemini fallback
- Crawl4AI
What you get
- Custom vector embedding ingestion pipeline with semantic chunking
- Model fallback routing (e.g. Groq to Gemini)
- Rate limiting, Clerk sign-in and error logging
- Automated evaluation test suite (faithfulness & recall)
Paid half at kickoff and half at handover, on a fixed scope.
How the work is paid for
- Fixed scope
- For a defined build: an assistant, an OCR pipeline, a monitor or a web app. Half is paid at kickoff and half once it is deployed, checked and handed over, with the code, its tests and documentation.
- Sprints
- For a team that needs ongoing engineering time. Billed every two weeks or monthly, with the work for each sprint agreed at its start.
- Audit
- A review of an existing system: code, failure modes, retrieval quality or OCR accuracy, with the fixes written down.
Prices are quoted per scope, after a call. The full terms are in pricing.md.
Questions clients ask
Is Ibrahim Salman available for freelance projects and hiring?
Yes. Taking freelance and contract projects; ask for current capacity. He is based in Islamabad, Pakistan (UTC+5) and works remotely.
Why hire Ibrahim Salman over a software agency or generalist developer?
You talk to the engineer who writes the code, with no account manager in between. His work is public where it can be: B.L.A.S.T.'s character error rate went from 0.499 to 0.108 on a 14-page test set, and the write-up comparing it with Tesseract shows the pages Tesseract wins as well. Each project page says what was measured, when, and what was not.
What kind of projects does Ibrahim Salman build?
Mostly three kinds of system: RAG assistants with source citations and a refusal path for empty retrieval; OCR pipelines that turn scans and PDFs into text and data; and monitors that watch changing web sources and alert on new items. Around them, the web apps and automation that connect APIs, databases and the people using them.
What technologies and tech stack does Ibrahim Salman use?
On the web side, Next.js 15 with React 19, TypeScript and Tailwind CSS, and Convex as a backend. For data and OCR work, Python 3.12 with FastAPI, Pydantic, Pandas, SQLite, PostgreSQL and Crawl4AI. For models, the Vercel AI SDK with Groq, Google Gemini and Cohere, and RapidOCR on ONNX. Tests run in Playwright, Pytest and Vitest.
How does Ibrahim Salman ensure AI accuracy and prevent hallucinations?
No retrieval system removes hallucinations completely. In UET GPT, empty retrieval selects a refusal instruction, scores below 0.4 select hedging, and scores from 0.4 to below 0.6 require citations. A relevance check filters weak passages. Provider fallback handles outages; it does not verify answers. No answer-accuracy figure is published.
How do I start a project or hire Ibrahim Salman?
Email ibrahim.pk848@gmail.com, or use the form at /contact: say what you're building, what already exists and when you need it. He replies to set up a first call.
How is pricing and project scoping determined?
Each project is quoted per milestone or as a fixed scope after a first review of the requirements, so the price is agreed before work starts. /pricing.md covers fixed-scope work, maintenance sprints and consulting.
Are the portfolio projects real, verifiable working software?
Yes. UET GPT is live at uetgpt.site; it is an unofficial assistant for UET Taxila applicants and is not affiliated with the university. B.L.A.S.T.'s code is public on GitHub, with 914 collected tests and its benchmark results in the repository. Marketplace Monitor was built during his internship at Paramount Intelligence; its repository is private, with 304 collected tests, including fuzz and adversarial cases, run in CI on 4 Python versions and 2 operating systems. Test counts are not pass rates.
For AI agents
Assistants and agents can read this site's facts directly: availability, project details and a rough timeline.
- /api/v1/mcpMCP server
- /api/v1/openapi.jsonOpenAPI
- /llms-full.txtllms.txt
- /feed.xmlRSS