Harshit Kashyap Sarma
Audio Technology / Full Stack / AI-ML
C
C++
Python
PyTorch
NumPy
librosaJavaScript
TypeScript
Next.js
Express.js
FastAPI
ZustandTanStack
Docker
MongoDB
MySQL
PostgreSQL
Redis
AWS Lambda
S3
Stripe
OAuth / PKCE
ABOUT
PROJECTS
OPEN SOURCE
EXPERIENCE
EDUCATION
BLOGS
About Me
I'm Harshit, an audio-tech and full-stack engineer who likes living where DSP meets machine learning. I build synths and VST plugins, ship multi-tenant backends, and train models for security and sound. Right now I'm heading to UPF Barcelona for a Master's in Sound & Music Computing.
Something - Beatles
Projects
Kick Generator A variational autoencoder (VAE) trained on kick-drum samples. It learns a compact latent space of percussive sounds and samples from it to synthesize brand-new kicks, morphing smoothly between punchy, boomy, and clicky textures.Tech Stack := Python, PyTorch, NumPy, librosa, matplotlib
Clueminati 2025 The official event app for CodeChefVIT's Clueminati 2025, a real-time, QR-driven treasure hunt shipped as an installable PWA. Players scan QR codes to unlock clues and watch a live leaderboard update. Built on Next.js with MongoDB and TanStack Query, it stayed responsive under hundreds of concurrent participants.Tech Stack := Next.js, TypeScript, MongoDB, TanStack
Open Source
ac-timbral-models MERGED #1PRIVATE+360-78#1 · Porting of the Sharpness Model
✓ Reviewed & approved by @ffont (MTG Lead / Freesound Director)·Oct 2026
Re-engineered the psychoacoustic sharpness model in native Essentia, reducing benchmark discrepancy from 24.21% to 0.36% with a 1.41× speedup.
- #Eliminated external dependencies (
scipy,soundfile,pyloudnorm) by porting directly to native Essentia algorithms (MonoLoader,LoudnessEBUR128,IIR). - #Extracted reusable ISO 532B (DIN 45631 / Zwicker) specific loudness module with precomputed static Butterworth & Chebyshev filter coefficients across 28 third-octave bands.
- #Reduced average error across benchmark suite from 24.21% down to 0.36% while running 1.41× faster than the original Scipy implementation.
- #Architected for zero-overhead translation into native Essentia C++ with zero external DSP dependencies.
Stack := Python, Audio DSP, Essentia, NumPy, C++
Private MTG repository (UPF research) — access restricted to organization collaborators
Freesound MERGED #2166PUBLIC+70-22#2166 · Fix 'Last Post' & Thread Counters on Spammer Cascade Deletion
✓ Reviewed & approved by @ffont (Freesound Lead)·Jul 2026
Resolved silent signal failures during spammer account cascade deletion and built a production management command to heal historical database corruption.
- #Fixed
post_deletesignal handler where cascade deletions raised silentProfile.DoesNotExistexceptions, halting thread and forum metadata repairs. - #Replaced profile decrements with queryset
.update()no-ops and clamped counters withGreatest(..., 0)to guarantee out-of-sync safety. - #Authored unit and regression tests covering both direct user deletion and soft profile purge execution paths.
- #Shipped
update_last_thread_forum_postsDjango management command to scan and repair historical corrupted forum rows in production.
Stack := Python, Django, PostgreSQL
Experience
KnowledgeVerse AI Software Development Engineer Intern · Bengaluru, IndiaMay 2025 – Aug 2025 | Dec 2025 – Present- #15× faster file/folder listing across the multi-tenant platform via a Redis caching layer with TTL-based invalidation.
- #Built and published the official Python SDK (kv-platform) to PyPI — auto-generated, strongly typed, with secure API-key auth.
- #Migrated the backend from sync to async (boto3 → aioboto3) across 18 files, removing event-loop blocking and speeding up report generation.
- #Shipped an end-to-end Stripe integration — dynamic plans, Payment Intents, subscription webhooks — with zero payment incidents.
- #Secured all auth flows with Google & Microsoft OAuth SSO (Authorization Code Flow + PKCE) across Next.js and FastAPI.
- #Built a serverless upload pipeline on AWS Lambda + S3, extended to Google Drive and OneDrive/SharePoint with encrypted uploads and retry handling.
Stack := Python, FastAPI, Next.js, Redis, AWS Lambda, S3, Stripe, OAuth / PKCE, aioboto3, asyncio, PyPI
- #Authored the “SynthQuest: Building a VST from Scratch” blog series on real-time audio DSP and the JUCE plugin pipeline.
- #Ran hands-on audio-programming workshops at local schools, teaching the fundamentals of sound synthesis and DSP.
- #Led the technical execution of Clueminati, CookOff, and DevSoc — flagship hackathons with 1000+ combined participants.
Stack := Next.js, TypeScript, MongoDB, C++, JUCE, Audio DSP
- #Oversaw technical operations of GraVITas '25, one of India's largest student-run fests with 40,000+ participants and 200+ events.
- #Led and mentored a team of developers building the portal and admin panel with role-based access for events, logistics, and R&R.
Stack := Next.js, TypeScript, TanStack, Axios
Education
Universitat Pompeu Fabra (UPF) Master's in Sound & Music Computing (SMC)Barcelona, Spain · 2026 – Present- #To Be Started
VIT Vellore Bachelor of Science in Computer ScienceVellore, India · 2023 – 2026Top Course Projects
›Open-Port Vulnerability Scanner (Machine Learning): Flags vulnerabilities on exposed, idle open ports, classifying risk with XGBoost and Random Forest models trained on port and service features.
›Deep Fingerprint (Capstone): Detects malicious HTTPS connections via a hybrid approach: a rule-based engine paired with a neural network trained on the full feature set that composes a JA3 fingerprint.
