Open to backend & ML rolesBSc Information Technology · APU / APIIT · Class of 2029

Sakshyam Shrestha

Backend Developer · ML Engineer

Recently shipped Aqua-Guard, a flood-risk model for Nepal's hydropower corridors, at JunctionX Kathmandu 2026.

I build the models and the services around them: PyTorch classifiers, FastAPI and Django APIs, and the React front ends that put them in front of people. Currently an undergraduate in Information Technology at APU.

RésuméLinkedIn

Works with

  • PyTorch
  • FastAPI
  • Django
  • React
  • PostgreSQL
  • Docker
Kathmandu, Nepal27.7250° N, 85.3200° EAqua-Guard · JunctionX Kathmandu 2026
[02]

About

Not a list of soft skills. Three claims, each one carrying a number.

Portrait of Sakshyam Shrestha
Kathmandu, Nepal

I am Sakshyam Shrestha — an undergraduate in Information Technology at APU who builds machine-learning systems that survive contact with real, messy data, and the APIs and interfaces that put them in front of the people who need them.

My focus is applied modelling with honest evaluation: chronological splits, station-relative normalisation, and metrics that survive an imbalanced target. The work below is the evidence.

  1. 01Aqua-Guard

    F1 0.91 · recall 0.83

    A flood-risk classifier for Nepali hydropower corridors, built in a 30-hour hackathon window at JunctionX Kathmandu 2026. I handled the model, the feature pipeline and the evaluation — scoring precision 1.00 and recall 0.83 on the flood class, which is only 4.2% of the data.

  2. 02Saral Kitab

    Offline-first

    Inventory and sales tracking for small retail, built to stay usable with no signal: a Django REST backend behind token auth, an Android client, and a stale-while-revalidate cache so the last known state is always readable.

  3. 03KaloKot

    9 heuristics

    Procurement corruption-risk analysis. Scores a tender across nine rule-based red-flag heuristics and answers follow-up questions grounded in the Constitution of Nepal 2015 and the Public Procurement Act 2063, citing the article and naming the oversight body.

[03]

Skills

Self-assessed against shipped work. Every value is defensible from a repository on this site.

AI / ML

Depth evidenced by shipped models, not coursework.

PythonPyTorchModel evaluationData pipelinesLLM & agentsAnomaly detection
  • Python5/5Every model, API and script behind this site.
  • PyTorch4/5Aqua-Guard — classifier written and trained by hand, no pretrained weights.
  • Model evaluation4/5Chronological splits, ROC-AUC and F1 instead of raw accuracy.
  • Data pipelines4/5Gauge discharge and weather CSVs from the Chisapani and Melamchi basins, fused into station-relative features.
  • LLM & agents3/5KaloKot's retrieval layer, Chronicle's interpreter and the X.A.R.V.I.S agent.
  • Anomaly detection3/5Aqua-Guard's anomaly service flags readings that fall outside the historical envelope.

Backend / Web

APIs, auth and the front end that consumes them.

FastAPIREST designReactDjango / DRFAuth & dataTooling
  • FastAPI4/5Aqua-Guard's /predict service and KaloKot's analysis API.
  • REST design4/5Documented contracts, real status codes, typed payloads.
  • React3/5Aqua-Guard's risk dashboard and the Saral Kitab app.
  • Django / DRF3/5Saral Kitab's token-authenticated backend.
  • Auth & data3/5Token auth, role-based authorisation, offline caching.
  • Tooling3/5Git, Netlify and Render deploys end to end.

Supporting tools

Used in shipped work
Typer
CLI entry points for Chronicle and CacheWay.
LangChain
X.A.R.V.I.S agent pipelines.
Scikit-learn / Pandas
House Price regression pipeline.
Streamlit
Model demo interfaces.
Git & GitHub
Atomic commits, branch flow, public repos.
ChromaDB
Constitution semantic search in KaloKot.
HTML / CSS
Layout and UI fundamentals.

Archived

Historical only
PHP / MySQL
LMS only — kept for reference, not current practice.
[04]

Projects

Selected builds, most recent first. Each entry leads with the result, then shows how it was built.

01JunctionX Kathmandu 2026Model, features & evaluation

Aqua-Guard

Flood-risk classification for Nepal's hydropower corridors

Classifies flood risk at river gauge stations so hydropower operators get an early warning instead of a post-hoc report. Rated-discharge and weather series for the Chisapani and Melamchi corridors are turned into station-relative features, a PyTorch network is trained on the flood class, and a FastAPI service serves the predictions to a React dashboard. Anomaly detection, compliance scoring and a Gemini-backed voice summary sit alongside the model. Built end to end inside a 30-hour hackathon window.

Precision 1.00 and recall 0.83 on the flood class — which is just 4.2% of the data, so raw accuracy is not the honest headline.

F1
0.91
Recall
0.83
Precision
1.00
Accuracy
~99%
  • Chronological split — the model never sees the future
  • Flood class: 4.2% of records
  • Python
  • PyTorch
  • FastAPI
  • React
  • TanStack
  • Scikit-learn
Live dashboardSource
Live dashboard · 10 stations
02

Saral Kitab

Inventory & sales tracking for small retail, built for patchy connectivity

An Android app that lets a shop owner record sales, returns and restocks from the counter without a stable connection. React and Capacitor on the front, a Django REST Framework backend behind token auth, and a stale-while-revalidate cache so the last known state stays readable when the network drops.

Turns a paper ledger into a live revenue, margin and inventory-value view — usable with no signal.

  • React
  • Capacitor
  • Django REST Framework
  • Token Auth
APISource
Inventory · transactions · analytics
03

KaloKot

Procurement corruption-risk analyzer & virtual lawyer

Upload a government tender PDF and KaloKot extracts the text, scores nine rule-based red-flag heuristics — suspicious timeline, tailored specifications, budget inflation, missing evaluation criteria, emergency procurement, conflict of interest, one-sided contract terms — and returns a traffic-light heatmap. The scoring engine runs with zero LLM calls. Ask a follow-up and a digital lawyer answers from the Constitution of Nepal 2015 and the Public Procurement Act 2063, cites the article, names the oversight body and drafts the complaint.

Turns a 90-page tender into a red-flag heatmap and a ready-to-file complaint letter in about a minute.

  • Grounded in: Constitution of Nepal 2015 · Public Procurement Act 2063
  • CIAA · PPMO · OAG referral paths
  • Python
  • FastAPI
  • Gradio
  • LLM
  • ChromaDB
  • React
Live appSource
In Chambers · counsel chat

Chronicle

04

Local-first developer intelligence

Point it at a Git repository and it scans commit history and Django migrations into structured observations, analyses those into findings, then answers questions about the project in plain language. Everything stays in a local SQLite file inside a .chronicle/ folder — no account, no cloud sync, no telemetry — and reruns are incremental, because each scanner and analyser records how far it got.

Published on PyPI as chronicle-devkit: a question about your own project history, answered without a byte leaving the machine.

  • Python
  • Typer
  • SQLite
  • Git
  • OpenAI / Gemini
On PyPISource

CacheWay

05

Caching proxy for repeated upstream reads

A FastAPI proxy that intercepts GET requests to a target origin, forwards anything it has not seen upstream, and answers repeats from an in-memory cache on a 60-second TTL. The origin and port are Typer CLI arguments rather than config files, so a different upstream is a different command, not an edit.

Repeat reads served from memory instead of the network.

  • Python
  • FastAPI
  • httpx
  • Typer
Source

X.A.R.V.I.S

06

Command-line AI agent

A LangChain agent that turns a natural-language request into structured system actions from the terminal — tool selection, planning and execution without a browser in the loop.

Task automation driven entirely from the shell, no GUI.

  • Python
  • LangChain
  • CLI
  • AI Agents
Source

LMS

07

Library management system

A multi-role PHP/MySQL application for cataloguing, issuing and tracking — the piece that taught me relational schema design, session state and role-based authorisation the hard way.

A working multi-user catalogue with enforced borrowing permissions.

  • PHP
  • MySQL
  • HTML
  • CSS
Live siteSource

House Price Prediction

08

End-to-end regression pipeline

Scikit-learn regression over real USA housing data, wrapped in a Pandas cleaning pipeline and shipped as an interactive Streamlit interface so a non-technical person can query it directly.

A trained model anyone can interrogate through a browser, not a notebook.

  • Python
  • Scikit-learn
  • Pandas
  • Streamlit
Source

AniTalk

09

Anime review platform

A front-end review and discovery site — my first pass at layout, filtering and UI organisation before I had a backend to hold it together.

Where the fundamentals clicked.

  • HTML
  • CSS
  • JavaScript
Source

Repositories marked source are private pending publication.

[05]

Education

2026 — 2029

BSc Information Technology · Asia Pacific University of Technology & Innovation (APU / APIIT) · Software development, databases, applied AI/ML

2024 — 2026
Higher Secondary — Science, Computer · Bagmati Secondary School & College, Kathmandu · GPA 3.53 / 4.00
2014 — 2024
School — Science stream · Bhakti International School, Kathmandu · Head Prefect · GPA 3.83 / 4.00
[06]

Certificates

A hackathon build, an applied AI workshop, and the courseware underneath. Kept deliberately small.

29–31 May 2026

featured

JunctionX Kathmandu 2026

Hackjunction Finland · SUMS Nepal · COGKNIT · Embassy of Finland

30-hour hackathon, participated as Hacker on team Astral. Challenge track powered by eSewa. Built and shipped Aqua-Guard in the window.

11–13 Jestha

secondary

AI/ML Workshop — 3 days

Consultify Inc. with Leapfrog Connect

Three-day applied AI/ML programme run at Don Bosco College, Kathmandu. Certificate of Participation.

Issued 17 September 2026

secondary

Python

freeCodeCamp

Self-paced Python and object-oriented programming, representing approximately 300 hours of work.

ID fcc-e61bd4cb-9e60-4a1a-acae-b217d5241146/python-v9

Verify

13 July 2026

minor

Python Django 101

Simplilearn SkillUp

Introductory Django online course — the starting point for the Saral Kitab backend.

ID 10465726

15 July 2026

minor

Learn Python Intermediate

Programiz

Intermediate Python — data structures and application structure.

ID 73128F3A79A4

16 July 2026

minor

Practical Git and GitHub

Programiz

Version control workflow — branching, history and collaboration.

ID 70F1359EAA78

[07]

Contact

Available for machine-learning and backend roles — internships, freelance ML work, or a junior full-time seat.

I reply to email fastest. If you have a dataset, a model that is not reaching its accuracy target, or a backend that needs building, tell me about it.

sakshyamxtha35@gmail.com