~/ankit $ ls ./projects --all
Things I've built.
Research, AI applications, and full-stack systems, each one shipped, not just started.
NDN Anomaly Detection: Interest Flooding & Cache Pollution
Research under Prof. Babu Ram Dhawadi: unsupervised detection of Interest
Flooding and Cache Pollution attacks in Named Data Networking. One Isolation
Forest detector, two independent tracks: live miniNDN/NFD emulation
(PIT-level features) and ns-3/ndnSIM simulation (face-level tracer counters).
100% attacker detection at ~1s latency with 0% held-out false positives across
tree, dumbbell, and DFN topologies. KS hypothesis testing and feature-ablation
isolated a data-quality issue, raising detection from 8.5% to 61% with no
model change. Paper draft in progress.
pythonisolation-forestminindnndnsimunsupervised-ml
ongoing research
ragchatbot
Open-source, pip-installable RAG SDK (v1.1.3 on PyPI) that turns a document
folder into a production-ready semantic-search chatbot API with zero ML setup.
A pluggable multi-LLM layer (Gemini, OpenAI, or local Ollama) behind one
interface, a ChromaDB vector store with MD5-based incremental indexing, token
streaming, and a 25-test pytest suite.
pythonragchromadbclipypi
English ↔ Nepali Neural Machine Translation
Encoder-decoder Transformer built from scratch in PyTorch (Vaswani et al.):
multi-head self/cross-attention, sinusoidal positional encoding, label smoothing,
and a warmup LR schedule, no high-level NMT libraries. A 35M-parameter model
trained on 1M sentence pairs filtered from Meta AI's 19.6M-pair NLLB corpus,
scoring 13.8 BLEU / 41.2 ChrF on the standard FLORES-200 benchmark.
SentencePiece BPE and length-bucketed batching cut ~80% padding overhead and epoch
time ~12x.
pytorchtransformernmtnepalisentencepiece
[04]
computer vision
Padel Game Analytics
Built during my AI/ML internship at Layman AI: a CV pipeline that analyses
padel match video in real time: YOLOv8 player/object detection, TrackNet
ball tracking, MediaPipe pose estimation, rule-based shot-type classification,
and annotated video output with per-frame CSV/JSON analytics.
yolov8tracknetmediapipeopencv
Kanun Sathi: AI Legal Assistant
RAG platform (team of 4) answering legal questions in Nepal with verifiable
page-level citations to source law (Constitution, criminal/family law).
PDF chunking, 768-dim Ollama embeddings, pgvector similarity search,
and Gemini 2.5 Flash with citation-enforcing prompts; async ingestion via
Redis/ARQ workers. React + TypeScript frontend, full stack in Docker Compose.
fastapiragpgvectorgeminireact
Nepali NLP Platform
Web platform showcasing NLP models for the Nepali language, spoken by
30M+ people yet underrepresented in NLP. React + Flask + Hugging Face
Transformers, Dockerised and cloud-deployed.
reactflaskhuggingfacedocker
Nagarik Suchana
Civic engagement app connecting citizens with local authorities: notices,
grievances, governance transparency. Auth, admin dashboard, English/Nepali
support, OCR via Tesseract.js, Docker-containerised services.
react-nativenode.jspostgresqldockerocr
Infant Cry Classification
End-to-end audio classifier for infant cries (hunger, pain, discomfort) from
uploaded or recorded audio. TensorFlow LSTM/ANN on MFCC + spectral features
from 1,000+ recordings, with SMOTE for class imbalance. Exported to
TFLite and served by a Flask API to a React Native (Expo) app (iOS/Android/Web)
with a consistent MFCC pipeline end to end.
tensorflowlstmtflitereact-nativeflask
MCP Multi-Server Chat
Chatbot wired to multiple Model Context Protocol servers: local math
tools, a remote expense tracker, and Manim animation generation, all through
a LangChain MCP client with a Streamlit UI.
mcplangchainstreamlitmanim