Sourav Chakraborty

Revision Oct 2026
Guwahati, Assam

AI backend engineer. I build FastAPI services that serve ML models, secure them, and deploy them.

Available for 2026 internships

Features

  • Two deployed AI APIs with live demos
  • ~91% deepfake-audio detection (Wav2Vec2, FastAPI)
  • JWT auth, CORS, OWASP Top 10 hardening
  • CI/CD with GitHub Actions
  • Open source: GSSoC 2026, OSCG
Figure 1. A speech signal. Overlay a synthetic one to see the kind of artifacts VoiceGuard looks for.
Peak −3.1 dB Duration 2.0 s Pitch ≈ 140 Hz

Your audio stays in your browser. To check if a clip is a deepfake, run it through VoiceGuard AI.

Description

What I build and where I study

I'm a second-year B.Tech CSE (AI & Data Science) student at Assam Down Town University in Guwahati, originally from Tripura. I work where ML inference meets backend engineering: designing REST APIs around models, adding authentication and hardening, and getting them deployed with CI/CD.

I'm deliberate about the gap between a demo and a production system. Right now that means moving my projects to async inference, containers and proper service layouts, and rebuilding one of them without AI-written code so I understand every line.

Typical application A client sends an HTTPS request to a FastAPI service protected by JWT and CORS. FastAPI passes the input to a model and returns a JSON response. The service is hosted on HuggingFace Spaces or Render, and the frontend on Vercel. Client FastAPI Model JSON response HTTPS Frontend on Vercel JWT · CORS Hosted on HF Spaces / Render Typical application A client sends an HTTPS request to a FastAPI service protected by JWT and CORS. FastAPI passes the input to a model and returns a JSON response. The service is hosted on HuggingFace Spaces or Render, and the frontend on Vercel. Client FastAPI Model JSON response HTTPS Frontend on Vercel JWT · CORS Hosted on HF Spaces / Render
Figure 2. Typical application. Both projects follow this pattern: a client calls a FastAPI service, which runs a model and returns JSON.
Degree
B.Tech CSE (AI & Data Science)
University
Assam Down Town University
Batch
2025–2029
CGPA
8.05

Application notes

Projects, how they're built, live demos

VoiceGuard AI

Deepfake audio detection API

Feb 2026

Cloned voices are hard to tell apart from real ones by ear. For the India AI Impact Buildathon by HCL GUVI, I built a REST API that takes an audio clip and returns a score and a label: natural or synthetic speech. A FastAPI service wraps Meta's Wav2Vec2-Large-XLSR-53 model from HuggingFace. Requests are protected with JWT auth and CORS rules, secrets stay in .env, and every change deploys through GitHub Actions.

Specifications
Detection accuracy~91%
ModelWav2Vec2-Large-XLSR-53
FrameworkFastAPI
AuthJWT, CORS
HostingHuggingFace Spaces
ContainersDocker (local builds)
CI/CDGitHub Actions
ContextIndia AI Impact Buildathon (HCL GUVI), 40,000+ participants
VoiceGuard AI block diagram An audio upload goes to a FastAPI service, then to the Wav2Vec2-Large-XLSR-53 model, which produces a score and label. A GitHub Actions box feeds deployment of the FastAPI service. Audio upload FastAPI Wav2Vec2-Large-XLSR-53 Score + label GitHub Actions deploy
Figure 3. VoiceGuard AI. An uploaded clip is scored by Wav2Vec2-Large-XLSR-53 behind FastAPI. GitHub Actions deploys the service.

Status: In progress: moving sync inference (run_in_executor) to a routers / schemas / services layout.

MediSense AI

Symptom analysis backend

Mar 2026

MediSense turns a plain-language description of symptoms into structured condition cards, each with a probability score and an urgency level for triage. A FastAPI service on HuggingFace Spaces calls Gemini 2.5 Flash and shapes the response. The Vercel frontend never talks to the model directly: it goes through a secure API bridge, and the backend is hardened against the OWASP Top 10.

Specifications
ModelGemini 2.5 Flash
FrameworkFastAPI
HostingHuggingFace Spaces (API), Vercel (frontend)
IntegrationSecure API bridge
AuthJWT, CORS
SecurityOWASP Top 10 hardening
OutputCondition cards with probability and urgency
MediSense AI block diagram A Vercel frontend calls a secure API bridge, which reaches a FastAPI service on HuggingFace Spaces. FastAPI calls Gemini 2.5 Flash, and the result comes back as condition cards with probability and urgency. Vercel frontend Secure API bridge FastAPI HuggingFace Spaces Gemini 2.5 Flash Condition cards probability + urgency
Figure 4. MediSense AI. The Vercel frontend reaches Gemini 2.5 Flash only through the FastAPI service, by way of a secure API bridge.

Status: In progress: rebuilding MediSense without AI-assisted code. AI explains, I implement.

AudioGuard Agent

Deepfake-audio risk checks for developer pipelines

GitAgent Hackathon

An agent that runs inside a developer pipeline, scores audio for deepfake risk, classifies how confident that score is, and writes an audit report. Built for the GitAgent Hackathon.

Specifications
InputAudio in a developer pipeline
OutputRisk score, confidence class, audit report
ContextGitAgent Hackathon
AudioGuard Agent pipeline Audio enters the developer pipeline, is risk scored, and produces an audit report. Audio in pipeline Risk scoring Audit report
Figure 5. AudioGuard Agent. Audio is risk scored and the result is written up as an audit report.

Also built: Firework Simulator v3

Interface

Ask about me the way you'd call an API

Endpoints
Try:
200 OK 38 ms 1.1 KB
{
  "name": "Sourav Chakraborty",
  "role": "AI backend engineer",
  "location": "Guwahati, Assam",
  "availability": "Available for 2026 internships"
}

Runs in your browser against this page's data. No server involved.

Characteristics

Each skill, rated, with the project that proves it

Skills, with proficiency rating and where each was used
Parameter Rating Test condition
PythonShippedVoiceGuard, MediSense
FastAPIShippedVoiceGuard, MediSense
JWT auth / CORSShippedVoiceGuard, MediSense
GitHub ActionsShippedVoiceGuard CI/CD
HuggingFaceShippedWav2Vec2 on Spaces
Gemini APIShippedMediSense
DockerProjectLocal builds
SQL, C++CourseworkB.Tech
MLflowBasicsRoadmap exercises
Async FastAPILearning6-week MLOps plan
Redis / CeleryLearning6-week MLOps plan
WebSocketsLearningNone yet
  • Shipped (used in a deployed project)
  • Project or coursework
  • Learning

Revision history

Experience and milestones in order

Experience and milestones, newest first
Date Change
Apr 2026–nowContributor, GirlScript Summer of Code 2026 (AI Agents and Open Source tracks)
Mar 2026–nowOpen Source Contributor, OSCG. Merged PR, Contributor Badge
Mar 2026Shipped MediSense AI
Feb 2026Shipped VoiceGuard AI at India AI Impact Buildathon (HCL GUVI)
2025Started B.Tech CSE (AI & DS), ADTU

Also a member of GDG On Campus ADTU.

Errata and roadmap

What I'm fixing and learning right now

Fixing

  • Moving VoiceGuard and MediSense to containerised, CI-driven workflows
  • Rebuilding MediSense independently

Learning

  • Async FastAPI (ASGI lifecycles)
  • Redis and Celery
  • WebSocket streaming
  • Database connection pooling
  • Multi-stage Docker builds
  • IBM AI/DS and Azure MLOps certifications

I write technical breakdowns on LinkedIn, where my network is 1,300+ people.

Last updated: October 2026

Ordering information

How to reach me

The fastest way to reach me is email.

Download resume