Skip to work

Open to Junior AI / platform roles

Data · AI · Infrastructure Engineer

I build systems that turn messy data into decisions, and keep them running.

Four disciplines, one job: move a model from a notebook to something a business can depend on at 3 a.m. I work end to end — feature pipelines and evaluation, inference services and agents, and the network and cloud they run on.

See the workDownload resume
Kathmandu, NP · Remote friendly@kritikashresthain/kritikashresthakritikashrestha216@gmail.com

DATA

Pipelines & analysis

SkillsFeature engineering·Experiment design·Forecasting·Causal analysis
StackPython·dbt·Airflow·DuckDB·Spark
FocusData contracts·Reproducible training sets

Hover or tap a node

Signals

Rolling totals, updated with each shipped project.

+4 this year

0

Projects shipped

production, not demos

stable

0

Years engineering

since 2019

+1 in Q2

0

Certifications

AWS · Cisco · Databricks

3 incidents

0.00%

Platform uptime

12-month rolling

Route taken

I did not plan this order. Each stage exposed the thing the previous one could not solve, and I followed it.

  1. 2017

    UI/UX design

    Where it started

    Designing dashboards taught me that a number nobody trusts is worse than no number at all.

    FigmaDesign systems

  2. 2019

    Data science

    Chasing the numbers

    Moved behind the chart to the pipeline. Learned that most model failures are data failures wearing a costume.

    PythonStatisticsdbt

  3. 2021

    Applied AI

    From models to products

    Shipped retrieval and agent systems, and built the evaluation harnesses that decided whether they were allowed to ship.

    PyTorchRAGEvals

  4. 2023

    Network & systems

    The floor everything stands on

    Inference latency turned out to be a routing problem. Went down a layer, and stayed.

    BGPCiliumZero-trust

  5. Next

    Self-operating platforms

    Where I am heading

    Systems that observe themselves, explain their own regressions, and hand engineers a diagnosis instead of a dashboard.

    AIOpsAutonomous remediation

Kathmandu, NP·UTC+05:45

Track record

Case studies

Each one is a system that had to survive contact with real traffic.

AI2025In production

Agentic AI Interview Platform

A screening pipeline that runs itself, end to end

An agent-driven hiring pipeline that handles intake through reporting without a coordinator in the loop, while keeping every decision auditable.

Problem

Recruiters spent 60% of their week on scheduling and first-round screens, and scoring drifted between interviewers with no record of why a candidate was rejected.

Approach

Split the pipeline into narrow agents with typed handoffs, each writing a structured artifact. Scoring runs against a rubric with held-out calibration sets, and any low-confidence decision escalates to a human instead of guessing.

Features

  • Typed handoffs between agents, so failures localise
  • Rubric-based scoring calibrated against human panels
  • Confidence threshold that escalates instead of guessing
  • Full decision transcript exportable for compliance

4d → 6h

Time to first screen

0.81 κ

Scoring agreement

-64%

Cost per candidate

Architecture

01

Intake

Parse CV, normalise to schema

02

Scheduling

Calendar solve across panels

03

Interview

Structured agent session

04

Evaluation

Rubric scoring + calibration

05

Reporting

Auditable decision record

Stack

Next.jsPythonLangGraphPostgrespgvectorTemporal
Infrastructure2024In production

SignalMesh

Network telemetry to root cause in under a minute

A telemetry pipeline that correlates flow data, BGP events and service metrics into a single ranked cause list during an incident.

Problem

During outages, three teams stared at three tools and argued about which layer was at fault. Mean time to restore sat near 40 minutes, most of it spent on attribution.

Approach

Normalise every source onto a shared topology graph, then score candidate causes by how well each explains the observed symptom set. Engineers get a ranked list with the evidence attached, not a wall of charts.

Features

  • Topology-aware correlation across network and application layers
  • Ranked causes with the evidence trail attached
  • Replay mode for post-incident review
  • Backpressure-safe ingest at 400k events per second

40m → 11m

Mean time to restore

400k/s

Ingest throughput

-52%

False page rate

Architecture

01

Collect

sFlow, BMP, OTel

02

Normalise

Map onto topology graph

03

Correlate

Symptom to candidate causes

04

Rank

Score by explanatory power

05

Notify

Page with evidence attached

Stack

GoClickHouseKafkaGrafanaTerraform
Data2023Open source

FeatureForge

One definition of a feature, for training and serving

A feature store layer that compiles a single declarative definition into both the training query and the online serving path.

Problem

Training and serving skew kept surfacing weeks after deploy, because the offline SQL and the online code were written by different people at different times.

Approach

Make the definition the only artifact. Compile it to dbt models for training and to a typed serving client, then diff the two paths nightly and fail loudly on drift.

Features

  • Single spec compiled to both training and serving paths
  • Nightly skew detection with a hard CI gate
  • Point-in-time correct backfills
  • Typed Python and TypeScript clients generated from the spec

11 → 0

Skew incidents

-73%

Backfill time

6

Teams adopted

Architecture

01

Define

02

Compile

03

Materialise

04

Verify

Stack

PythondbtRedisSnowflakePydantic
Infrastructure2024Internal

EdgeGuard

Zero-trust routing for inference traffic

An identity-aware routing layer that keeps tenant inference traffic isolated without adding a proxy hop to the hot path.

Problem

Multi-tenant model serving needed hard isolation, but every off-the-shelf option added 30-60ms to requests already fighting a latency budget.

Approach

Push policy into eBPF at the node level so enforcement happens in kernel space, and keep the control plane out of the request path entirely.

Features

  • Kernel-space policy enforcement, no sidecar hop
  • Per-tenant flow audit with 30-day retention
  • Policy changes propagate in under two seconds
  • Fails closed on control plane loss

<1ms

Added latency

1.4s

Policy propagation

40+

Tenants isolated

Architecture

01

Identity

02

Policy

03

Enforce

04

Audit

Stack

CiliumeBPFGoKubernetesSPIFFE

Toolkit

Depth reflects what I have run in production, not what I have read about.

Data Science

Turning raw events into training sets that hold up.

  • Python95
  • dbt88
  • Airflow84
  • Spark76
  • DuckDB82

AI

Retrieval, agents, and the evals that keep them honest.

  • PyTorch88
  • LangGraph85
  • vLLM90
  • pgvector83
  • Weights & Biases78

Networking

Routing, telemetry, and finding the layer that actually broke.

  • BGP86
  • Cilium / eBPF80
  • Ansible88
  • Wireshark84
  • NetBox79

Development

Shipping the interfaces and services around the models.

  • TypeScript90
  • Next.js87
  • Go78
  • Postgres85
  • Terraform86

Design

Interfaces for operators who are already having a bad day.

  • Figma82
  • Design systems80
  • Data visualisation84
  • Accessibility77

Signature

Four domains, one loop. Select a domain to see what it carries and what it hands off.

Capabilities

  • Experiment design
  • Forecasting
  • Feature engineering
  • Causal inference

Tools

PythondbtDuckDBAirflow

Hands off

Clean, point-in-time-correct features

to AI Engineering

Foundations

M.Sc. Computer Science, Data Systems

Tribhuvan University

2017 — 2019Kathmandu, NPDistinction

Coursework

Distributed SystemsStatistical LearningComputer NetworksDatabase InternalsOptimisationInformation Retrieval

Achievements

  • Thesis on point-in-time correctness in feature pipelines
  • Teaching assistant for the undergraduate networks lab

Certifications

  • AWS Solutions Architect — Professional

    Amazon Web Services · 2024

  • CCNP Enterprise

    Cisco · 2022

  • Certified Kubernetes Administrator

    CNCF · 2023

  • Databricks ML Professional

    Databricks · 2023

  • HashiCorp Terraform Associate

    HashiCorp · 2022

Get in touch

Tell me what is breaking

Hiring, consulting, or a system that will not behave — send the details and I will reply within two working days.

Replies within 2 working days

GitHub@kritikashresthaLinkedInin/kritikashresthaEmailkritikashrestha216@gmail.com

Fill in all three fields to continue.