I'm a Computer Science student at Delhi Technological University (DTU) interested in the intersection of:
🤖 Artificial Intelligence · ⚙️ Systems · 🛠️ Developer Tools · 🔐 Security · 🚀 Product Engineering
I enjoy taking vague ideas, weird technical problems and ambitious prototypes and turning them into working software.
My current rabbit holes:
LLMs · RAG · AI Agents · Local Inference · Backend Systems · Developer Experience · Security
Build it. Break it. Understand it. Build it better.
|
DTU CSE B.Tech Computer Science Class of 2029 |
9.35 Current CGPA |
19 Public Repositories |
Making Indian legal information easier to understand.
Legal.ai is a citizen-first AI legal assistant designed around a retrieval-first architecture.
It combines legal retrieval, semantic search, document understanding, multilingual interaction, voice input and real-time response streaming.
- 🧠 Retrieval-Augmented Generation
- 🔎 Semantic vector search
- 🕸️ LangGraph orchestration
- 📄 Legal document processing
- 🎙️ Voice interaction
- 🌍 Multilingual interaction
- ⚡ Real-time streaming
- 📚 Grounded responses
🥉 3rd Place — Build With Gemma Hackathon
A MySQL client that doesn't look like it escaped from 1998.
PY DBMS is a terminal-first database client focused on developer experience, security and usability.
- 🔑 Argon2 authentication
- ⚙️ Multi-layer configuration
- 📊 CSV / JSON exports
- 🧾 Multi-line SQL
- 🎨 Rich terminal UI
- 🚩 Query-level flags
- 🧪 Automated tests
- 📦 PyPI packaging
Because developer tools are products too.
RESUME → CLASSIFICATION → FEATURE EXTRACTION → SHORTLISTING
An end-to-end ML pipeline exploring what happens when individually reasonable ML components are connected into a real decision system.
- 📚 Multi-class classification across 24 roles
- 🔤 TF-IDF
- 🧬 Feature engineering
- 🌲 Random Forest
- 📈 Logistic Regression
- 🧪 End-to-end evaluation
- 🔍 Failure analysis
The most interesting result wasn't the accuracy.
It was discovering that:
Good models can still produce a bad system.
One configuration ended up rejecting 99%+ of candidates, exposing how distribution mismatch and chained model assumptions can dominate end-to-end behaviour.
That failure taught me more than a perfect benchmark ever could.
What if your code optimizer didn't need to upload your code?
OPTIMA experiments with local AI-powered code optimization.
Instead of sending source code to a remote API, inference happens locally.
SOURCE CODE
⬇️
STATIC ANALYSIS
⬇️
PROMPT BUILDER
⬇️
LOCAL LLM
⬇️
OPTIMIZED CODE
- 🖥️ Local inference
- 🧵 Browser workers
- 📡 Streaming generation
- 🧠 Context limits
- 📦 Large-input chunking
- ⏱️ Model latency
- 🛟 Fallback strategies
- 🔒 Privacy-first architecture
On-device AI gives you privacy — and then hands you every other problem.
A deliberately simple personal finance widget.
No giant dashboard.
No 17 charts.
No spreadsheet disguised as an app.
Just:
Built with:
TypeScript · React · Nothing ecosystem
A 90% accurate model can still produce a terrible system.
The first version asks:
"Does it work?"
The tenth version asks:
"Where does this code belong?"
Local AI.
Limited compute.
Latency.
Memory.
Security.
Privacy.
These aren't annoyances.
They're the engineering problem.
The bugs, weird outputs, performance cliffs and unexpected failures usually teach me more than successful demos.
If a tool is technically impressive but painful to use...
it's still painful to use.
|
|
|
|
Computer Science student. Builder. AI/ML enthusiast. Systems nerd. Occasional over-engineer.
I like:
→ building developer tools → breaking my own systems → understanding why they broke → experimenting with AI → turning prototypes into products
Build something useful. Make it fast. Make it clean. Make it impossible to stop thinking about.
I enjoy building software that goes beyond a classroom submission.
My repositories explore:
AI/ML · Developer Tools · Databases · Security · Cryptography · CLI Apps · Web Apps · Experimental Systems
If you find something useful:
⭐ Star it · 🐛 Break it · 💡 Suggest something · 🔀 Fork it
