CSE student at GITAM University. I build systems that work in real environments — not demos, not notebooks, not university assignments.

The common thread across everything I build: connecting an intelligence layer to something physical or operational. Rocket telemetry. Underwater autonomy. Clinical AI. Production e-commerce. The code either works under real constraints, or it doesn't work.

📬 bc833498@gmail.com  ·  LinkedIn  ·  GitHub

Languages — Python, JavaScript, SQL, Java, C
Backend — FastAPI, REST API design, JWT auth, WebSockets, async I/O
ML / AI — PyTorch, Scikit-Learn, TF-IDF, CNNs, Grad-CAM, SHAP, MLflow, DVC
Autonomous Systems — ROS2, Kalman filtering, sensor fusion, real-time telemetry
Frontend — React, TypeScript, Vite, Chrome Extension APIs
Infrastructure — PostgreSQL, MongoDB, Docker, Prometheus, Grafana, Cloudinary

Python · FastAPI · WebSocket · Kalman Filter · PostgreSQL · React · TypeScript — 2026

Real-time ground control station for the IN-SPACe Model Rocketry Competition 2026, built for GITAM University's CAN-7USAT team targeting 1000 m AGL.

The backend decodes 46-byte binary telemetry packets from a Teensy 4.1 flight computer transmitted over a 900 MHz XBee radio link, runs Kalman filter sensor fusion, maintains a 6-state flight state machine, and broadcasts live data to a React dashboard over WebSocket.

Numbers that matter:

• End-to-end latency: < 5 ms (target was 15 ms)
• Packet decode time: 0.5 ms (target was 2 ms)
• WebSocket broadcast: 1 ms (target was 5 ms)
• Packet loss: 0% across all 23/23 passing tests

Packet structure: 46 bytes — sync byte, timestamp, flight state, altitude, velocity, quaternion (W/X/Y/Z), GPS lat/long, XOR checksum
State machine: PRE-FLIGHT → BOOST → COAST → APOGEE → DESCENT → LANDED
Stack detail: FastAPI + Uvicorn + asyncpg + NumPy/SciPy (Kalman) · React 18 + TypeScript + Zustand + uPlot (real-time charts) + Three.js

The challenge wasn't writing the code — it was writing code that works once, correctly, on launch day, with a rocket already on the pad.

→ Repository

Python · FastAPI · ROS2 · Docker · asyncio — 2026

Full autonomous control stack for an underwater vehicle. No internet, no margin for error. Every component had to work offline and degrade gracefully.

The system runs a deterministic observe → decide → act agent loop with 4-DOF thruster control and a safety monitor that enforces hard limits on depth, battery, temperature, and obstacle proximity before any decision executes.

• Offline LLM decision engine (llama.cpp / OpenAI-compatible API) with rules-based fallback when the model is unavailable or slow
• Agent loop p95 latency: ≈ 0.84 ms
• Append-only JSONL telemetry logging for full mission replay and post-mission analysis
• Structured telemetry covers every state transition, decision, and sensor reading

Designing the fallback logic was harder than building the main path. What does the vehicle do when the LLM times out at 20 m depth? That question drove the architecture.

→ Repository

PyTorch · FastAPI · Grad-CAM · SHAP · Docker · MLflow · DVC · Prometheus · Grafana — 2025–2026

Started at the AI Day Hackathon (GITAM × Kodryx AI, March 2026). Extended into a production-deployed clinical decision-support platform.

The system combines medical image analysis and structured clinical data (age, BMI, diabetes duration) to estimate ulcer risk — and critically, it shows why it made that call.

• Custom CNN in PyTorch for image classification
• Grad-CAM heatmaps highlight which image regions the model attends to
• SHAP values quantify how each clinical feature influenced the prediction
• Patient timeline tracking for ulcer progression monitoring
• Full MLOps stack: MLflow experiment tracking, DVC dataset versioning, Prometheus metrics, Grafana dashboards
• 5 production health-check endpoints (/health, /health/ping, /health/ready, /health/live, /health/status)
• JWT auth, RBAC, Cloudinary image storage, Docker Compose
• Live: diabetic-ulcer-ai-system.vercel.app

Black-box predictions are a liability in clinical settings. The explainability layer isn't a nice-to-have — it's the entire point. A clinician needs to understand, challenge, and override the system.

→ Repository

FastAPI · Chrome Extension · TF-IDF · Logistic Regression · LLM reasoning — 2025–2026

Started at HackXplore (IEEE-SSIT Student Chapter, VJIT, April 2026). A browser extension that scores news content for credibility as you browse — one click, instant result.

• Logistic Regression on 10,000+ labeled samples, 90% validation accuracy with TF-IDF vectorization
• Model loaded into memory at startup for sub-second response
• LLM reasoning layer generates contextual verdict explanations beyond the raw confidence score
• Chrome extension extracts DOM content and transmits it to the FastAPI backend via CORS-configured endpoints

The interesting design problem wasn't the model — it was latency. The pipeline had to feel instant. That forced early decisions about model loading, memory layout, and what work happens at startup vs. at request time.

→ Extension  ·  → Backend

FastAPI · PostgreSQL · SQLAlchemy · React · JWT · RBAC — 2025

Full-stack platform for real hospital workflows. The goal was to model the domain correctly, not just build CRUD endpoints.

• Relational PostgreSQL schema with ORM mapping — patient records, appointments, inventory, all designed around how the data actually relates
• Strict RBAC: Admin, Doctor, and Receptionist roles hitting entirely separate data partitions — enforced at the query level, not just in route handlers
• ACID-compliant transaction blocks for anything that touches multiple tables
• Decoupled React frontend and FastAPI backend, clean REST boundaries throughout

The hard part was the domain model. If the schema is wrong, no amount of application logic fixes it. I spent more time on the database design than on the API.

→ Repository

FastAPI · PostgreSQL · JWT · Razorpay — 2024–Present · Production, live

Backend for a real printing business. Not a demo. In use.

• 15+ REST endpoints, JWT auth, bcrypt hashing, multi-role access (User / Admin / Business)
• Razorpay payment integration
• 6-state order lifecycle (pending → confirmed → in-production → ready → dispatched → delivered) across 10 relational tables with ACID transactions

→ Repository

AI Intern — SmartED Innovations  ·  Feb 2026 – Present
Designing ML workflows in Python/Scikit-Learn, building FastAPI integrations for model inference pipelines

AI/ML Intern — Elevate Labs  ·  Apr 2026 – May 2026
Applied ML: preprocessing, training, evaluation, deployment-oriented backend integrations

MLOps in practice — model serving, versioning, monitoring, drift detection. Advanced backend patterns — connection pooling, async task queues, caching. Distributed systems basics. Deeper XAI for domains where the explanation matters as much as the prediction.

I learn fastest when something real is at stake. Most of what I know came from a deadline, a broken system, or a constraint that ruled out the obvious solution.