AKSHAT CHAUHAN

Backend Systems Engineer.

Systems built for scale, failure, and the messy realities of production.

IndiaRust · Go · CDistributed Systems3 yrs

Building real systems

Systems built under real-world constraints.

Not demos. Production-oriented systems shaped by tradeoffs between reliability, performance, scalability, correctness, and operational complexity.

Backend EngineeringDistributed SystemsAI/MLInfrastructure

Flagship System

SmartQueue

Adaptive ML Scheduler

v2.1ActiveProduction

Production-grade distributed task scheduler with adaptive priority queues, worker orchestration, LSTM runtime prediction, and Kubernetes-native autoscaling.

Multi-tenant

Predictive Scheduling

Kubernetes Native

Horizontal Autoscaling

Production Ready

Mode

Runtime

K8S

Region

PROD-EU

Scale

1 → 20

IsolationMulti-tenant queue

PredictionLSTM runtime

GatewayFastAPI + Prometheus

DeliveryGitHub Actions CI/CD

PythonFastAPIPostgreSQLDockerKubernetesPrometheus

Production Ready

GitHub

Active

Observability

Security · Rust · Tokio

EchoTrap

Async TCP honeypot built on Tokio. Captures and fingerprints attacker behavior across 10k concurrent connections with sub-50ms response latency. Emits structured logs for downstream SIEM integration.

RustTokioTCPAsync

Active

Observability

Systems · Rust · ptrace

SysRift

ptrace-based syscall recorder and deterministic replayer, ported from C to Rust. Captures full syscall boundary state and replays execution faithfully — useful for debugging non-deterministic production failures.

RustCptraceLinux

View all on GitHub

Operating real systems

Building systems under real-world constraints.

Experiences shaped by deployment constraints, production tradeoffs, and the responsibility to make systems work when reliability matters.

01.0 / Experience

AI Engineer

GeeksforGeeks KIIT

Operating inside applied ML systems where model accuracy, inference latency, and deployment constraints directly affected usability and system reliability.

Feb 2025 – PresentKIIT UniversityActive

PythonPyTorchOpenCVFlaskJetson NanoTensorRT

lip-read / deployment-pipeline

live inference

Inference pipeline

−35%

inference latency after TensorRT optimization

Architecture

Input video

30 fps

→

CNN encoder

7.8 ms

→

LSTM state

128 MB

→

Beam decoder

98.2%

→

Prediction

edge

CNN-LSTM pipeline quantized with TensorRT and deployed to Jetson Nano under a 128 MB memory budget. Cold-start guards prevent partial-state serving on resource-constrained restarts.

Deployment

confidence98.2%

runtime4 ms p95

targetJetson Nano

throughput412 req/min

servingFlask / edge

Change log

02.14CNN-LSTM pipeline promoted to edge validation.

02.18TensorRT pass deployed — latency reduced 35%.

03.04Image-to-LaTeX decoder added to inference surface.

02.0 / Experience

AI Engineer

IoT Lab, KIIT

Worked on applied NLP systems embedded in larger data and serving pipelines, where model behavior under noisy inputs and low-compute environments mattered more than offline metrics.

Jan 2025 – PresentKIIT UniversityActive

PythonTensorFlowPyTorchFlaskBERTLIME

news-classification / nlp-pipeline

evaluation

Classification pipeline

+15%

precision gain via active learning — no model growth

Architecture

Raw articles

100k+

→

BERT classifier

bert-base

→

Active learner

+15% precision

→

LIME explainer

feature weights

BERT classifier trained on 100k+ articles with an active learning loop that surfaces uncertain samples for selective re-labeling. LIME traces provide category-level explanation without additional inference cost.

Accuracy matrix

runtime11 ms

batch32

environmentedge

buildbert-exp.09

Change log

01.20BERT classifier indexed against 100k+ news articles.

02.03Active learning loop raised precision 15% without model growth.

02.11LIME traces enabled for category-level explanation review.

03.0 / Experience

Web Development Lead

AISOC

Led and operated a production-facing web platform where frontend decisions, backend APIs, and infrastructure constraints directly impacted reliability under event-scale load.

Sep 2024 – PresentKIIT UniversityActive

Next.jsTailwind CSSNode.jsMongoDBVercelCI/CD

aisoc / event-platform

deployed

Service architecture

250+

hackathon registrations processed under event-scale load

Architecture

Next.js

Vercel / edge

→

Node API

REST / server

→

Async workers

job queue

→

MongoDB

primary store

Next.js frontend with a Node API layer backed by an async job queue for evaluation processing. Async queue offloading reduced submission processing time by 40%, keeping p95 latency below 50 ms under burst registration traffic.

Production

uptime99.9%

p95 latency43 ms

queue depth8 jobs

regionap-south-1

releasev3.2.1

CI/CD

01 / checkout8 s

02 / lint43 s

03 / build1 m 12 s

04 / deployprod

Change log

09.18Next.js redesign shipped behind performance budgets.

10.06Registration flow handled 250+ hackathon submissions.

10.19Async evaluation queue reduced processing time by 40%.

03.0 / Competitive Programming

Algorithmic systems under contest constraints.

LeetCode + CodeChef

Competitive programming as an operating discipline: ranking progression, algorithmic consistency, and decision-making under hard time and correctness constraints.

2024 - PresentActiveRanking UpTop 15%

DSAGraphsDPGreedyBinary SearchContests

01LeetCode Analytics

ActiveKnight

Rating

1,772

+184

Rank

Knight

Top 9%

Solved

211

71%

Contests

Weekly

Rating progressionRanking Up

02CodeChef Analytics

Active4 Star

Rating

1,765

+211

Rank

4 Star

Top 12%

Solved

198

68%

Contests

Starters

Rating progressionConsistent

Contests

tracked

Problems Solved

409

combined

Best Rank

Top 9%

platform peak

Standing

Top 15%

global

Active Years

2024+

current

Tech Stack

Systems built with deliberate tools

Technologies I use to build, ship, and operate backend systems with an eye on reliability, performance, and maintainability.

Systems & Low-Level Programming

Performance, memory, concurrency

Rust

Cloud, Infra & Tooling

Deployment & operations

Docker

Kubernetes

AWS

Linux

Git

AI / ML (Applied)

Engineering-focused ML

PyTorch

OpenCV

10 technologies3 categories

I design software assuming I'll be the one paged when it fails.