Computer Science graduate (AI/ML) building AI applications across computer vision, speech processing, and NLP. I like turning models into working products: inference pipelines, APIs, and real-time systems that run on real hardware.
- 🎙️ Building an offline multilingual voice AI agent (Bengali, Hindi, English)
- 🧬 Building ThoughtField, a multi-agent social simulation engine
- 🏭 Shipping real-time CV systems with YOLOv8 and OpenCV at my internship
- 🔬 Contributor to open-source AI frameworks (Aden Hive, 8.5k+ ⭐)
| 🎯 99.3% | 🔎 99% | 🧠 4 GB → 1.2 GB | ⚡ ~0.8 s |
|---|---|---|---|
| Brain tumor MRI classification accuracy | Number plate OCR accuracy on test images | LLM memory after 4-bit ONNX quantization | Offline voice agent response time |
timeline
title Sourav Pati
2022 : Started B.Tech in CSE (AI/ML) at UEM
2025 : Oracle Certified AI Foundations Associate
: Oracle Certified Generative AI Professional
2026 : Graduated, built Voice AI Agent, ThoughtField and Brain Tumor Detection
: Joined Beas Consultancy as AI/ML Intern
Aug 2026 – Present
| What I built | |
|---|---|
| 🎙️ | Offline multilingual voice AI agent: speech recognition, language identification, local LLM inference, and text-to-speech |
| 🚚 | Vehicle number plate recognition: detects trucks, crops plate regions, extracts text with PaddleOCR at 99% accuracy on all test images |
| 🏭 | Real-time factory intrusion detection on RTSP CCTV streams with YOLOv8 + OpenCV: multi-threaded stream reading, auto-reconnect, IoU-based person tracking, zone-overlap checks for angled cameras |
| 🚨 | Live zone editor, dwell-time alarm with siren, and automatic snapshot + event logging (entry time, exit time, duration) to JSON |
| 🏷️ | Image annotation and labeled data preparation for an object detection project |
A locally running speech-to-speech agent supporting Bengali, Hindi, and English, combining multilingual speech recognition, language identification, and neural speech synthesis. Runs fully on CPU with ONNX Runtime and was tested offline.
flowchart LR
A["🎤 Speech input"] --> B["Language identification"]
B --> C["ASR<br/>Bengali · Hindi · English"]
C --> D{"Tier 1<br/>Lexical match"}
D -- hit --> H["🔊 Neural TTS reply"]
D -- miss --> E{"Tier 2<br/>ChromaDB semantic search"}
E -- hit --> H
E -- miss --> F["Tier 3<br/>Local 4-bit LLM"]
F --> H
F -. unanswered .-> G["☎️ Customer-care escalation"]
classDef neon fill:#0D1117,stroke:#00F5FF,color:#ffffff,stroke-width:2px
classDef warn fill:#0D1117,stroke:#FF2E97,color:#ffffff,stroke-width:2px
class A,B,C,D,E,F,H neon
class G warn
- ⚡ ~0.8 s for dataset-matched queries, 2-3 s for generated answers
- 🧠 Quantized the local LLM to 4-bit ONNX: memory down from ~4 GB to 1.2 GB
An AI-powered multi-agent social simulation platform based on the Stanford Generative Agents architecture, where autonomous agents interact and drive prediction-based simulations.
flowchart TB
UI["🖥️ Next.js frontend<br/>live simulation canvas"] <-->|"HTTP + WebSocket"| API["⚡ FastAPI backend<br/>REST + WebSocket"]
API --> R[("Redis<br/>pub/sub + state")]
API --> C[("ChromaDB<br/>agent memory vectors")]
API --> O["🧠 OpenAI API<br/>ticks · reflection · report"]
classDef neon fill:#0D1117,stroke:#00F5FF,color:#ffffff,stroke-width:2px
classDef data fill:#0D1117,stroke:#9D4EDD,color:#ffffff,stroke-width:2px
classDef ai fill:#0D1117,stroke:#FF2E97,color:#ffffff,stroke-width:2px
class UI,API neon
class R,C data
class O ai
- 🧠 Agent memory and semantic retrieval with ChromaDB, real-time workflows with Redis
- 📡 REST APIs plus WebSocket communication for live agent interactions
An end-to-end deep learning web app that classifies brain MRI scans into Glioma, Meningioma, and Pituitary tumors using a fine-tuned ResNet50 with transfer learning, reaching ~99.3% accuracy on the Figshare Brain Tumor MRI dataset.
flowchart LR
A["🩻 MRI scan"] --> B["OpenCV preprocessing<br/>+ data augmentation"]
B --> C["ResNet50<br/>transfer learning"]
C --> D["Glioma"]
C --> E["Meningioma"]
C --> F["Pituitary"]
C --> G["🌐 Flask web app"]
classDef neon fill:#0D1117,stroke:#00F5FF,color:#ffffff,stroke-width:2px
classDef out fill:#0D1117,stroke:#FF2E97,color:#ffffff,stroke-width:2px
class A,B,C,G neon
class D,E,F out
| Project | What it does | Tech |
|---|---|---|
| 🕸️ AI Agents: Aden Hive | Web page summarizer and resume analyzer agents on a self-evolving agent framework | Python, Hive, GraphSpec, LLM |
| 🎯 SHL Assessment Recommendation Engine | AI-powered system that recommends assessments based on user needs | Python, Streamlit |
| 👤 Face Identification Model | Attendance tracker using face recognition | Python, Computer Vision, API |
B.Tech in Computer Science and Engineering (AI/ML) · University of Engineering and Management · 2022 – 2026 · CGPA 7.28/10
Contributor to the Aden Hive Framework (YC-backed, 8.5k+ GitHub stars), working with GraphSpec / NodeSpec multi-agent architecture. Also into content writing.