PK
Open to work Backend · Full stack · ML infrastructure Available to start immediately
28.0836° N, 80.6081° W · Melbourne, Florida

Piyush
Kasle

Software Development Engineer

01 / About

The work around the model that keeps it running.

I'm a software engineer with a Master's in Software Engineering from Florida Tech and a year of backend experience at Acaira Technologies. I build APIs, databases and the infrastructure that keeps ML models running, and I care about tests, observability and clear tradeoffs.

Available to start immediately.

02 / Stack

From the API down to the server it runs on.

Languages

  • Python
  • Java
  • JavaScript
  • SQL

Backend & web

  • FastAPI
  • Flask
  • Express.js
  • Node.js
  • React

Databases & caching

  • PostgreSQL
  • MySQL
  • Redis

ML & computer vision

  • PyTorch
  • Hugging Face
  • scikit-learn
  • LightGBM
  • pandas
  • OpenCV
  • YOLOv8

Cloud & DevOps

  • AWS EC2
  • AWS S3
  • Docker
  • GitHub Actions
  • Linux

APIs & integrations

  • Google Meet API
  • Mailgun
  • Razorpay
  • JWT

Testing & tools

  • pytest
  • Postman
  • Git
  • Streamlit
  • Plotly

Embedded

  • C/C++
  • Arduino
  • STM32

03 / Projects

Built end to end, tradeoffs written down.

P1

mental-health-mlops

Case studyFall 2025Master's capstone

The ML side of a university health center's student check in app, built in my advisor's lab as my Master's capstone while the app was live with students. A fine tuned DistilBERT model sorts each message into one of six categories and is served by FastAPI on AWS EC2. Every message, prediction and confidence score is logged to PostgreSQL, a three tier drift monitor watches the inputs, and GitHub Actions tests, builds and deploys it.

  • DistilBERT
  • FastAPI
  • PostgreSQL
  • Docker
  • AWS EC2 + S3
  • GitHub Actions
  • Streamlit
F1 across 6 categories
81.9%
recall, highest risk class
74.1%
inference latency
98 ms
hosting, AWS Free Tier
$0/mo

Decision log

D1DistilBERT, after the simple baselines

I benchmarked Logistic Regression, Random Forest, LightGBM and LightGBM with SMOTE before reaching for a transformer. SMOTE resampling didn't help and class weighting worked better, but the simple models still weren't enough. Among transformers I picked DistilBERT for deployability: about 40% smaller and 60% faster than BERT while keeping about 97% of its performance.

chose · DistilBERT over · BERT and classic baselines
D2t3.small, not a micro instance

The model, the FastAPI service and the Streamlit dashboard need about 2 GB of RAM together, more than a micro instance has. t3.small is the smallest that fits, and the deployment still costs $0 a month on the AWS Free Tier. The model lives in S3 and is pulled in at Docker build time instead of sitting in git.

chose · t3.small over · t2.micro class
D3A bigger labeled batch before retraining

After the counselor's corrections, retraining on a dozen examples would have been quick and mostly noise. I asked her to label a larger batch from the app's logs instead, to use both for retraining and for testing the new drift signals.

chose · larger labeled batch over · retraining on 12 corrections
P2

security-shift-manager

Fall 2024Solo projectrebuilt from coursework

A three role Flask web app for scheduling and running a campus security team. Managers build and publish weekly rosters across 8 patrol locations and 3 shifts, guards see their duties and request leave or overtime, and an admin sets the pay rules behind payroll.

  • Flask
  • SQLAlchemy
  • MySQL
  • PyJWT
  • Jinja2
  • Bootstrap
  • Fair rosters by design. Each slot goes to the guard with the fewest duties this month, approved leave is respected, nobody gets two duties in a day, and a slot that can't be filled is recorded as uncovered instead of quietly producing an invalid roster.
  • Leave cancellation with an authority boundary. A pending request cancels at once; an approved one goes back to the manager for a second decision, tracked as its own flag so the leave still reads Approved until it's settled.
  • Double booking ruled out by the database. A composite unique key on duty, date and shift, with any clash shown as a clear message instead of a server error.
  • Payroll from the month's duties. Regular, night and overtime pay plus bonus, minus leave fines, regenerated in place rather than duplicated and released only after the month ends.
P3

smart-parking-system

Jun 2023 to May 2024Final year projectteam of 2

My undergraduate final year project, built with one teammate; I led the backend and the computer vision. A YOLOv8 and OpenCV service watches the live campus surveillance feed and tracks which slots are free in real time, even flagging a motorbike or bicycle parked in a car space. A separate Flask API runs accounts, vehicles and bookings by vehicle type, with JWT auth and Razorpay payments.

  • YOLOv8
  • OpenCV
  • Flask
  • PostgreSQL
  • JWT
  • Razorpay
slot detection accuracy
98%
concurrent users, load test
100
response time at that load
<200 ms
  • No double bookings. PostgreSQL row level locking on the slot being booked.
  • Validated in stages. Recorded footage first, then live against the real campus feed in the parking lot.
  • Two independent processes. The detection service and the booking API run separately.
P4

robocon

  • video ↗

CompetitionDec 2021 to May 2023Team Robocon CRCE

Two seasons on Team Robocon CRCE, my college's team in DD Robocon, India's national round of ABU Robocon. I went from writing the drive code to leading the computer science side and driving the robot in matches.

  • embedded C++
  • Arduino
  • STM32
  • PCB design
  • holonomic drive
All India Rank, 2022
AIR 21
All India Rank, 2023
AIR 23
integration, down from 1.5 days
~5 h

2022 · Lagori

Software lead

  • Owned the drive code for both robots and led software: me and three first year students, six weeks before the qualifying round.
  • Made every subsystem a plug in unit (mechanism, electronics, C++ module and self test) that had to pass a bench rig before going on a robot. Integration fell from about a day and a half to about 5 hours.
  • Worked across electronics and mechanical too: soldering, PCB design and wiring. Pit crew at the event.

2023 · Casting flowers over Angkor Wat

Computer Science Head

  • Led the computer science side: circuit design, the drive algorithm and the robot control code.
  • Drove in matches. That season's robots: a 30 kg Mecanum drive robot with flywheel shooters for poles up to about 10.5 m away, and a lighter omni wheel robot for the nearer, higher poles.
  • Designed Obstacle Race, the team's recruitment competition: the game, the scoring, the penalties and the arena.
photo of you at the competition
photo (optional)
photo (optional)

04 / Experience

From Mumbai to Melbourne, Florida.

  1. Jul 2020 to May 2024 Mumbai, India

    B.E., Electronics and Computer Science

    University of Mumbai · Fr. Conceicao Rodrigues College of Engineering

    Where I spent two seasons on Team Robocon CRCE and built the smart parking system as my final year project.

  2. Jun 2023 to May 2024 Mumbai, India

    Software Developer

    Acaira Technologies · Mumbai startup

    Backend for a school app serving about 1,100 students that had run on paper registers, WhatsApp groups and Excel. Express.js, MySQL, React and Redis.

    • Built the fee module backend, then proposed and built automatic receipts: PDFs in the school's format with gapless numbering, emailed from an idempotent queue, 4,300+ receipts and notices a year.
    • Added cheque tracking so a receipt only issues after a cheque clears: about 15 bounced cheques caught with zero receipts sent for them, and cheque handling cut by about 40%.
    • Took attendance straight from Google Meet, marking a student present at 75% of class time, which ended roll call and gave teachers back 5+ hours a week.
    • Cut the 12 KPI analytics dashboard from 3.2 s to 180 ms median with query rewrites, a covering index and Redis caching.
  3. Aug 2024 to May 2026 Melbourne, Florida

    MS, Software Engineering

    Florida Institute of Technology

    Graduated May 2026. Capstone: the MLOps pipeline above, built for the university health center. Coursework included Advanced Software Engineering, Advanced Software Testing, Software Architecture and Design, and Advanced Database Systems.

  4. Spring 2026 Melbourne, Florida

    Teaching Assistant, Database Systems

    Florida Institute of Technology

    A graduate course of about 28 students. Graded assignments and exams with written feedback aimed at the specific mistake, helped refine how the rubrics scored partial credit and common errors, and held weekly office hours.

05 / Contact

Open to new grad software engineering roles in backend, full stack and ML infrastructure. Available to start immediately.