Debarghya Saha
INTRO

Hi, I'm Debarghya.Applied AI Engineer & Data Scientist.

Building voice agents, RAG pipelines, and fast inference that hold up in production.

View Projects →Download Resume
ACHIEVEMENTS
Smart India Hackathon 2024 Winner
Led team to 1st place, Ministry of Education Problem 1742, top 0.3% of 88,221 teams.
FINALISTAI4Andhra Police Hackathon
WINNERSmart India Hackathon 2023 Internal
LISTEN
Hear the difference: raw vs. fine-tuned Hinglish TTS
Production voice agent: sub-200 ms latency
clip 0:12 · Hinglish
Debarghya Saha
About

The engineer behind the models.

I got into AI by trying to make machines talk the way people around me actually talk switching between Hindi and English mid-sentence. That turned into building voice agents, squeezing models onto CPUs, and caring a lot about latency.

I like small teams, fast feedback loops, and shipping things people use. Outside work: anime and manga, competitive hackathons, and taking apart old Game Boy ROMs.

Removed GPU dependency entirely.

BEFORE
CUDA serversGPU required · PyTorch
THE MOVE

Converted PyTorch models to ONNX Runtime and re-architected CUDA workflows into a quantized CPU-only pipeline, validated on Raspberry Pi class hardware.

PyTorch→ONNX→INT-quantized→CPU
AFTER
Raspberry PiCPU-only · edge class hardware

Systems in production.

Inference Engineering184 ms
LIVE CALL · on-device LLM
Caller: Mera order abhi tak deliver nahi hua, can you check?
Agent: Bilkul, let me check. Aapka order kal tak pahunch jayega.
STT → LLM → FastSpeech2 (Hinglish)3 GB artifact
Voice Agent

Real-time multilingual agent on an on-device LLM. Fine-tune training time by 32% and shipped as a 3 GB deployable artifact.

↗
SELF-HOSTED RAG QA
> What does the 2023 policy say about refunds?
milvus.search(HNSW, top_k=5) · 500,000+ docs
[1] policy_2023.pdf · p.14[2] faq_refunds.md[3] terms_v4.html · §7
Refunds are allowed within 30 days of purchase if [1]… ▍streaming · vLLM
RagProd

HNSW indexing in Milvus over 500,000+ documents, with streaming answers from vLLM.

↗
REINFORCEMENT LEARNING
$ python train.py --env PokemonRed-v0 --algo PPO
ep 1200 | battles 400 | win_rate 0.41
ep 2400 | battles 800 | win_rate 0.63
ep 3600 | battles 1200 | win_rate 0.77
ep 4800 | battles 1600 | win_rate 0.82 ✓
custom Gym env · ROM disassembly · PPO + A2C
Pokémon Red
RL AgentCustom Gym environment built from ROM disassembly, trained with PPO and A2C. 82% win rate.Read case study ↗

The stack.

Inference & Serving
ONNX RuntimevLLMllama.cppFastAPIQuantization
ML / AI
PyTorchHugging Face TransformersLangChainscikit-learnStable Baselines3LLM Fine-Tuning
Data & Storage
PostgreSQLMongoDBMilvusPineconeApache BeamApache KafkaAirbytePandas
MLOps / Cloud
DockerKubernetesTerraformGitHub ActionsMLflowPrometheusGrafanaAWSGCP
Speech
FastSpeech2TTS/STTSSMLFreeSWITCHAsterisk
EXPERIENCE
Junior Data ScientistEthical Den (Hivyr.ai)
Oct 2025 – Present
AI InternEthical Den (Hivyr.ai)
Jul 2025 – Sep 2025
EDUCATION
B.Tech, Computer Science and Business Systems
Meghnad Saha Institute of Technology2022 – 2026
Sudhir Memorial Institute, Doltala
2020 – 2022
Calcutta Airport English High School
2007 – 2020
CONTACTGot a model that needs to run faster, cheaper, or on a CPU?debarghyasren@gmail.com
GitHubLinkedInTwitter / XBlog · HashnodeDownload Resume ↓Kolkata, India