Projects
VOICE CLONE
- Model Architecture: AI4Bharat IndicF5
- Backend Engine: FastAPI & Uvicorn
- Cloud GPU Runtimes: Modal A10G Workers
- Audio Synthesis: Librosa / Pydub / Torch
- Reference Sample: 3–10s Clip + Transcript
- Long-Form Capacity: Up to 4,500 words chunked
01 / About Project
A production-grade full-stack voice cloning service that ingests a short 3–10 second Telugu audio clip alongside its transcript, producing studio-fidelity speech for any submitted text script. Powered by AI4Bharat's IndicF5 neural model and hosted via persistent Modal volume mounts with asynchronous status polling.
02 / Problem It Solves
Standard inference architectures timeout on HTTP endpoints and exhaust GPU VRAM during lengthy audio rendering. Voice Clone Studio decouples user request loops via background GPU queues, automatically truncating multi-thousand-word inputs into ~50-word syntactic chunks and stitching them with calculated acoustic silence padding.
01010111 01000001 01010110 01000101 00100000 01000111 01000101 01010010 01000001 01010100 01001001 01001111 01001110 00100000 [INDICF5 GPU WORKER AUDIO BUFFER] ~~~~~_~~~~_~~~~~_~~~~~_~~~~~_~~~~~_~~~~_~~~~_~~~~~_~~~~~_~~~~~_~~~~~_~~~~_~~~~_~~~~~_~~~~~_
Voice Clone Studio
[INSERT PROJECT IMAGE HERE]
Click this interactive exhibit to launch the deployed Telugu synthesis interface.
03 / Challenges Faced & Engineering Breakthroughs
Published IndicF5 weights were wrapped in torch.compile(), prefixing keys with _orig_mod.. HuggingFace silently failed key matching, yielding randomly initialized silence. Solved by writing load_checkpoint_with_key_remap() to strip prefixes directly from checkpoint shards.
Developed debug_repeat_call.py to execute four controlled model invocations on Modal, separating text-content discrepancies from instance weight initialization corruption without guessing.
Resolved persistent "Job Not Found" tracking failures across ephemeral container instances by migrating localized dictionaries into universally shared modal.Dict and persistent volumes.
Engineered automated standard-deviation calculations on generated speech buffers to intercept zero-variance silent failures instantly and emit user-facing exceptions.
TPCE PORTAL
- Core Frontend: HTML5 / Modern CSS3
- Logic Controller: JavaScript (ES6+)
- Hosting Server: Render Cloud
- Target Audience: Students & Faculty
01 / About Project
A centralized institutional portal engineered for TPCE college. Replaces outdated informational flyers with an instant, mobile-optimized digital hub showcasing academic departments, collegiate events, and student resources.
02 / Challenges Faced
Achieving silky-smooth 60 FPS visual interactions and responsive layouts without introducing heavy CSS framework bloat. Solved by constructing light native CSS architecture and hardware-accelerated micro-animations.
TOKEN NORMALIZER
- Language: Python 3 Core
- Linguistic Lib: NLTK Toolkit
- Pattern Engine: Compiled Regex Automata
- Target Use: ML Embedding Prep
01 / About & Problem Solved
Raw scraped conversational datasets contain irregular syntactic noise and unstandardized abbreviations that severely degrade downstream text vector embeddings and transformer training efficiency. This tool establishes a reproducible, deterministic token standard prior to modeling.
02 / Challenges Faced
Eliminating catastrophic exponential regex backtracking across megabyte-scale text arrays by utilizing compiled deterministic finite automata and lazy Python memory evaluation generators.
VISIOEXPLAIN
- Platform: Chrome Extension
- API Layer: Tabs & DOM Injection
- Core Logic: JavaScript (ES6+)
- Achievement: Hackathon Winner
01 / About & Problem Solved
Dense digital web layouts present severe cognitive hurdles for users experiencing informational fatigue or visual processing disabilities. VisioExplain dynamically inspects active DOM trees to generate intuitive structural summaries directly over live browser tabs.
02 / Challenges Faced
Coordinating asynchronous messages between background service workers and isolated content scripts while traversing strict Cross-Origin sandbox constraints without inducing visual page reflow lag.
Who I Am
B.Tech CSE (AI & ML) · TPCE '28
Sharmista Janapati
I am an AI/ML Software Engineer and university student specializing in deep acoustic neural models, distributed GPU worker architectures, and expressive modern web development. My focus lies at the crossroads of rigorous computational algorithms and intuitive, breathtaking human-computer interaction design.
What Tech Stack I Use
Foundational languages, cloud worker runtimes, and deployment protocols
Contact Me
Initiate connection for engineering collaborations & AI developmental ventures
Academic Curriculum Vitae
Download my official documentation and degree achievement records.