AshfakNawshadAI/ML
AI undergraduate at Moratuwa, learning by building. I work on document understanding, retrieval and reasoning systems — and I'm stubborn about making a model's answer traceable back to its source.
I'm drawn to the unglamorous half of applied AI: making a model's output provable.
University of Moratuwa · Second year
I'm an Artificial Intelligence undergraduate at the University of Moratuwa, focused on document AI, retrieval-augmented systems and neuro-symbolic reasoning. Most of my work starts from the same question — how do you stop a language model from inventing a number?
My flagship project, SME-GPT, answers it with a four-stage pipeline: OCR post-correction that is forbidden from altering digits, layout-aware serialization that carries provenance through every transform, and a planner-executor split where the LLM writes the plan and a deterministic runtime does the arithmetic.
Alongside all of that, I've tutored mathematics to 105+ students across the Sri Lankan and Edexcel syllabi. Explaining calculus to a room is the fastest way to find out whether you actually understand it.
SME-GPT
Explainable financial-document understanding for Sinhala and English. SMEs upload invoices, POs, receipts and delivery notes; the system extracts structured data and answers natural-language questions with grounded, provenance-backed, arithmetic-safe results.
Language models are fluent liars about numbers. For a small business reconciling invoices, a hallucinated total is worse than no answer at all — and Sri Lankan documents add bilingual Sinhala/English OCR noise on top. The system had to be auditable end to end: every figure traceable to a pixel region in a source document.
Semantic OCR post-correction
A correction pass repairs recognition errors in text while being structurally forbidden from altering digits, so cleanup can never silently rewrite a financial value.
Layout-aware serialization
Spatial document structure is serialized into a linear representation that carries provenance through every transform — each token keeps a pointer back to its region on the page.
Neuro-symbolic PAL question answering
The LLM plans; it never computes. It emits a program, and a deterministic executor runs the arithmetic — which removes the entire class of math hallucinations rather than mitigating it.
Multi-tenant relationship index
Documents are linked across a tenant's corpus so questions can span an invoice, its purchase order and its delivery note in a single grounded answer.
Second-year industry-based research project, in active development with a three-person team. Every answer the system returns cites the document region it came from.
- AI Search Algorithm VisualizerA browser-based tool that animates eight classical AI search algorithms — BFS, DFS, DLS, IDS, UCS, Bidirectional, Greedy and A* — step by step over graphs you build yourself, with export to PNG, GIF, PDF and SVG.
- Marine Learning HubAn AI study companion for marine-science students: log daily learnings, manage modules, and auto-generate flashcards and image-based specimen quizzes from your own notes and uploads.
- CodefolioA browser extension that turns your GitHub repositories into a polished resume. Connect an account, pick repos, refine the sections in a CV editor, and export in Modern or Academic templates.
- LocalPDFA privacy-first PDF toolkit — merge, split, remove, reorder, extract — running entirely on your machine with zero uploads. One engine, three front ends: CLI, local web UI and packaged desktop app.
- Flight Tracker AI AgentA conversational agent for real-time flight tracking, airport information and route search. The model decides which tools to invoke, then composes the results into an answer.
- Scholarship Eligibility AdvisorA rule-based expert system that matches students to scholarships over GPA, income, nationality and activities — deductive reasoning rather than statistical learning, with every decision explainable as a proof tree.
The tools are replaceable. The mathematics underneath is not.
Applied mathematics is the job. Pure mathematics is the reason the job is interesting — calculus, linear algebra and statistics are the working vocabulary behind every model I build.
Model training, retrieval systems and LLM orchestration.
Turning pages of unstructured paper into verifiable structured data.
Python first; the rest as the problem demands.
Shipping models behind APIs people can actually use.
Day-to-day infrastructure and deployment.
Where the model meets the physical world.
Mathematics, taught the way I wish it had been explained to me.
Alongside the AI work I tutor mathematics across the Sri Lankan local syllabus and Edexcel AS & A Level. It keeps the fundamentals sharp — you cannot hand-wave a derivation in front of a student who is about to sit the exam.
- —Grade 9–11 Mathematics
- —G.C.E. O/L Mathematics
- —G.C.E. O/L ICT
- —G.C.E. A/L ICT
- —Mathematics
- —Computer Science
- —Pure Mathematics 1–4
- —Further Pure FP1 · FP2 · FP3
- —Statistics S1 · S2
- —Mechanics M1 · M2
“Sir is a dedicated and supportive Mathematics teacher who goes the extra mile for his students. When I was struggling academically, he personally guided me and taught me simple, effective methods to improve. Within three weeks, I gained confidence and was able to progress towards achieving a good pass.”