engineer that ships.
I'm Praansu — BSc (Hons) Computing with AI student in Kathmandu, interning as a Junior AI Developer at Aviyaan Tech since March 2026. I work in PyTorch, RAG pipelines, and agent tool-calling, and I build the full-stack products around them. Every project below exists because I wanted to learn one specific thing properly.
Scroll — the work is below
FIG. 01 — Selected work
Seven builds, each chasing one question — agent loops without frameworks, RAG that deletes properly, classifiers measured honestly. No tutorials, no copy-paste.
Confusion matrix + per-class accuracy, not just 90%
ResNet18 transfer learning classifying bus/car/motorcycle/truck at 90% accuracy (400 images). Training set labeled with my own Tkinter annotation tool. Includes confusion matrix visualization and FastAPI inference API.
PyTorch · ResNet18 · FastAPI · Docker
Time-series feature engineering + model comparison, not just accuracy
Time-series forecasting for parking occupancy — synthetic data generation, feature engineering (lag/rolling/cyclical), Linear/RF/XGBoost comparison. EDA notebook with heatmaps, autocorrelation, seasonal decomposition.
Python · pandas · scikit-learn · XGBoost · statsmodels
ESP32 + FastAPI + real-time dashboard
IoT smart parking system — ESP32 firmware, FastAPI backend, real-time web dashboard with AI chatbot and booking system.
ESP32 · FastAPI · REST polling · IoT
Reliability ≠ model size; code-specialization beats raw scale
Research: evaluating 9 models (1B–9B) as autonomous agents across 31 capability tasks + 14 reliability tasks. Qwen 2.5 Coder 7B leads at 85% composite reliability.
Python · Ollama · Statistical Analysis · LaTeX
FIG. 02 — Build notes
Short notes, no tutorials. The question I started with, what I did, and the one result that stuck.
NOTE-01Small Agent Reliability
Do small models work as tool agents?
Scored 9 open-weight models, 1B to 9B, across 31 capability tasks and 14 reliability tasks — all run locally through Ollama, so the study reproduces on a laptop.
Reliability is not model size. Code-specialised training beats raw scale.
QWEN 2.5 CODER 7B — 85% COMPOSITE
NOTE-02AI Research Agent
Can an agent loop stay readable with no framework?
Hand-rolled loop in plain Python: route each question to document search, web search, or both — and stream every tool call to the browser over SSE as it happens.
Transparency is a feature. Watching the loop work builds more trust than any answer.
~120 LINES OF PLAIN PYTHON
NOTE-03Vehicle Image Classifier
Can 400 images make an honest classifier?
ResNet-18 transfer learning over four vehicle classes, served behind FastAPI and containerised with Docker.
One headline number lies. Per-class accuracy plus a confusion matrix shows which class the model is guessing at.
90% OVERALL — REPORTED PER-CLASS
FIG. 03 — Capabilities
Depth in ML and LLM systems, enough full-stack to ship them alone. Every entry here was earned by a project that needed it.
FIG. 04 — Background
Entry 01
Aviyaan Tech
Interning as a junior AI developer: building and experimenting with ML models — ensemble methods for traffic-count estimation, plus road-damage segmentation and classification — from data prep through training and evaluation.
Entry 02
Islington College, Kathmandu
Undergraduate degree focused on AI and ML. Coursework and self-directed builds in PyTorch, RAG pipelines, agent tool-calling, and full-stack deployment.
Entry 03
Global School of Science
Higher secondary education in the Science stream with Computer Science. Foundation in physics, chemistry, mathematics, and programming.
Entry 04
Personal & demo projects
Learning in public: RAG loops, agent harnesses, vision models, and a full-stack demo storefront (EcoVerda). No client work yet — available for freelance.
FIG. 05 — Services
Production-ready AI/ML systems and full-stack products — contract work and collaborations worth doing.
JOB-01OPEN
RAG pipeline development — FastAPI + ChromaDB + Groq/OpenAI, streaming SSE, document CRUD
JOB-02OPEN
Custom agent loops — No-framework agent design, tool calling, RAG + web search, transparent fallbacks
JOB-03OPEN
ML model deployment — PyTorch models → FastAPI + Docker, GPU/CPU optimization
JOB-04OPEN
Full-stack AI products — Next.js + TypeScript + Prisma + Stripe, production-ready
JOB-05OPEN
ML model training & deployment — PyTorch, transfer learning (ResNet, MobileNet, EfficientNet), FastAPI + Docker
JOB-06OPEN
Computer vision pipelines — Image classification, object detection, data augmentation, model optimization
JOB-07OPEN
MLOps — Model versioning, CI/CD for ML, Docker containerization, GPU/CPU optimization
JOB-08OPEN
Full-stack e-commerce development — Next.js 15/16 + TypeScript + Prisma + Stripe, production-ready
FIG. 06 — Contact
Freelance projects, full-time roles, or a good argument about RAG chunking — my inbox is open. I reply within a day or two.