hassan@habib:~$ whoami
CS undergrad @ Habib University, Karachi — NLP researcher, automation developer.
Two peer-reviewed papers before graduating. Ships tools, not demos.
hassan@habib:~$ cat /etc/motd
My portfolio is a chatbot trained on my résumé. It answers questions about me
so this README doesn't have to. This README exists anyway. I like redundancy —
every system I build has a fallback, including this introduction.
hassan@habib:~$ uptime
currently: Developer @ Office Automation Services · shipping LLM pipelineshassan09070/
├── 🔬 research/
│ ├── semeval2026 ······· DimABSA — VA regression, 6 languages · 🥈 2nd (Tatar), 6th (Russian)
│ └── clef-labs-2026 ···· FinMMEval — structure-aware RAG + MoE · 🌍 10th on the global leaderboard
├── 🛠 products/
│ ├── portfolio ········· a portfolio you argue with instead of scroll (beta)
│ ├── sequence-ai ······· board game AI — minimax thinking 23k states ahead of you
│ ├── khaata360 ········· WhatsApp bookkeeper — "spent 500 on food" → ledger entry
│ └── takhleeq ·········· 6-DOF robot arm math, twice (Python + TS), agreeing to 4 decimals
├── 🧗 grind/
│ └── neetcode250 ······· ████████░░░░░░░░░░░░░░░░░░░░ 69/250 — all 250 before 2027
└── 🌱 upstream/
└── sugarlabs ········· 5 merged PRs into software running in classrooms worldwide
📄 Habib University at SemEval-2026 Task 3: A Pipeline Approach for Dimensional Aspect-Based Sentiment Analysis
Proc. of the 20th International Workshop on Semantic Evaluation (SemEval-2026), pp. 3449–3459 — ACL
Sentiment as a continuous valence–arousal space, not three labels. A four-stage multilingual pipeline (mDeBERTa extraction → pairing → category → regression) across 6 languages, 4 domains — including a Double-[NULL]-token trick for implicit aspects. 🥈 2nd place — Tatar · 6th — Russian →
code
📄 HU_LLM_Fin @ FinMMEval 2026 Task 2: A Structure-Aware Hybrid RAG Pipeline utilizing Mixture-of-Experts
CLEF 2026 Working Notes — Jena, Germany
Standard RAG chunking shreds financial tables — a row without its header is a number without a year. Row-aware chunking + partitioned two-stage retrieval + prompt-caged Llama-4-Scout MoE, over SEC filings and news in five languages. 🌍 10th globally (GPT-4o scores 9.79% on this benchmark) →
code
Merged into Sugar Labs — LLM context management for the AI reflection feature in Music Blocks, educational software used in classrooms worldwide:
| PR | shipped | |
|---|---|---|
merged ✓ |
musicblocks#5991 | Sliding-window conversation memory with summarization |
merged ✓ |
musicblocks#6120 | Reflection widget UX — centering + refresh feedback |
merged ✓ |
reflection_fastapi#5 | Sliding window, server side |
merged ✓ |
reflection_fastapi#3 | Cut Gemini API cost via thinking_budget |
merged ✓ |
reflection_fastapi#4 | Docs for the above |
closed ✗ |
zulip#37279 | Emoji indicators for GitHub webhook notifications |
closed ✗ |
gemini-cli#19633 | Track deleted & corrupted sessions |
The ✗ rows stay listed on purpose. Shots taken > shots hidden.
|
The portfolio that answers back. An AI twin grounded strictly in my résumé — if the LLM dies, a local keyword engine takes over, tagged offline mode. The site cannot break. I checked. Repeatedly.
|
Bookkeeping by text message. WhatsApp in, ledger out — natural language and receipt photos parsed by vision LLMs. Built to a full SE process: SRS, SDS, Jira sprints. University course, industrial discipline.
|
🦾 TakhleeqMotor math for a real robot arm — the Dawlance × Habib Summer Tehqiq research program, building a 6-DOF arm from scratch in a country where nobody manufactures one. Same physics implemented twice (Python GUI + web); a parity test keeps both honest to 4 decimal places.
|
A board game that fights back. The Sequence board game with opponents that think ahead — Minimax with Alpha-Beta pruning, 23,000+ game states per move at expert. Pure-Python engine with zero dependencies, full stack layered on top.
|
panels and snake are hand-rolled SVG, regenerated nightly at 03:00 UTC by a workflow in this repo — no third-party stat service left to 503
hassan@habib:~$ ./contact --any-channel
email hassan.shahzad.info@gmail.com
linkedin linkedin.com/in/muhammad-hassan-shahzad-hu27
fastest ask my portfolio bot — it knows my schedule better than I do
hassan@habib:~$ exit
logout
Connection to hassan closed. (but the automations are still running)