Software engineer and AI/ML researcher building at the intersection of intelligent systems, mobile engineering, and data-driven product development. Pursuing a B.Tech in Artificial Intelligence & Data Science at IIIT Raichur (Class of 2028), I design and ship production-grade solutions across the full engineering stack — from TFLite on-device inference pipelines and deep computer vision architectures to cross-platform Flutter applications and FastAPI microservices.
My work spans research-grade ML engineering (computer vision, generative AI, adaptive signal processing) and applied software product development, with an unwavering emphasis on code correctness, system design clarity, and measurable real-world impact. I take pride in writing software that ships — not just prototypes that demo. Equally comfortable in a compiler, a tensor graph, and a boardroom.
|
Engineering Pillars
|
Open To
|
| Domain | Proficiency | Details |
|---|---|---|
| Computer Vision | ████████░░ Advanced |
DreamID · ArcFace · BiSeNet · CodeFormer · 3DDFA-V2 · OpenCV |
| On-Device ML | ████████░░ Advanced |
TFLite crash detection · Accelerometer fusion · Edge inference optimisation |
| Generative AI Pipelines | ███████░░░ Proficient |
14-stage face swap · Poisson blending · Cross-attention temporal conditioning |
| Adaptive Signal Processing | ███████░░░ Proficient |
LMS algorithm · Perceptron architectures · Filter convergence analysis |
| Data Mining & Analytics | ███████░░░ Proficient |
Apriori · FP-Growth · Clustering · Web mining · Feature engineering |
| MLOps & Deployment | ██████░░░░ Intermediate |
FastAPI inference servers · Model quantisation · Pipeline orchestration |
| NLP & LLM Engineering | ██████░░░░ Intermediate |
Prompt engineering · RAG patterns · Agentic workflow design |
🛡️ CrashGuard — Intelligent Road Safety Platform
A production-grade Flutter application combining on-device TFLite crash detection with real-time emergency response orchestration. Built under hackathon constraints, CrashGuard demonstrates end-to-end ML product engineering — from sensor fusion and accelerometer pipeline correction to offline-first UX and global emergency dispatch — and received recognition for technical depth and real-world impact.
| Attribute | Details |
|---|---|
| Stack | Flutter · Dart · TFLite · FastAPI · Python · OSM Overpass API · flutter_map · Hive |
| Scale | Android & iOS · Offline-first architecture · Global emergency numbers database |
| Performance | Corrected 10× accelerometer sampling rate mismatch · On-device inference < 150ms |
| Security | Local-first Hive storage · Zero PII transmission · Encrypted local boxes |
| Impact | Hackathon recognition · Full crash detection-to-emergency-dispatch pipeline |
| Repository | github.com/Mandeep15686/CrashGuard |
Three-module architecture — Flutter mobile client, FastAPI inference server, TFLite ML pipeline. The accelerometer fusion module resolves a critical sampling rate mismatch, achieving reliable binary crash classification on-device. OSM Overpass API queries surface nearby hospitals and police in real time; an offline first-aid guide and global emergency numbers database persist via Hive for zero-connectivity resilience. A full static code review identified 18 bugs prior to final submission, all resolved before deployment.
🎭 DreamFace — 14-Stage Generative Face Swap Pipeline
An industry-grade computer vision pipeline implementing a complete face swap system using state-of-the-art deep learning components. The architecture prioritises photorealism, identity preservation, and temporal consistency — engineered for production deployment, not research demos. A v2 redesign was executed following a rigorous architectural critique cycle, improving identity fidelity across diverse lighting conditions.
| Attribute | Details |
|---|---|
| Stack | Python · PyTorch · DreamID · 3DDFA-V2 · ArcFace · BiSeNet · CodeFormer |
| Scale | 14-stage pipeline · Full training suite · Modular component separation |
| Performance | Poisson blending for seamless compositing · Cross-attention temporal conditioning |
| Security | Consent-first design · Local processing · Zero cloud data exfiltration |
| Impact | Research-grade v2 architecture · Improved identity retention under lighting variation |
| Repository | github.com/Mandeep15686/faceswap-pro |
Stages 1–4: face detection, alignment, and landmark extraction via 3DDFA-V2. Stages 5–8: identity embedding with ArcFace and semantic segmentation via BiSeNet. Stages 9–11: swap synthesis through DreamID with cross-attention temporal conditioning. Stages 12–14: perceptual enhancement via CodeFormer, Poisson blending for seamless integration, and temporal smoothing for video consistency. A complete training suite with modular component separation enables reproducible experimentation and isolated ablation studies.
⚡ Developer Portfolio — Glassmorphism Engineering Showcase
A premium personal developer portfolio built with a glassmorphism aesthetic and a procedural neural network canvas animation system. Engineered entirely in vanilla HTML/CSS/JS to demonstrate frontend depth without framework overhead — every animation, interaction, and data visualisation is hand-crafted, from the RAF-based neural canvas to the Chart.js engineering metrics dashboard.
| Attribute | Details |
|---|---|
| Stack | HTML5 · CSS3 · Vanilla JavaScript · Chart.js · Canvas API |
| Scale | Single-page application · Multiple interactive modules · Fully responsive |
| Performance | RAF-based neural canvas · CSS GPU acceleration · Sub-2s LCP |
| Security | Static hosting · Zero server-side attack surface · CSP headers |
| Impact | Live portfolio · Skill tabs · Project modals · Animated engineering metrics |
| Repository | github.com/Mandeep15686/portfolio |
Core architecture: procedural neural network canvas animation using the Canvas API with requestAnimationFrame and node-edge graph simulation; glassmorphism card system built on CSS backdrop-filter with layered blur; dynamic skill-tab switching with animated underlines; project modal overlays with keyboard trap and ARIA compliance; Chart.js-powered engineering metrics block with animated reveal on scroll entry. Zero external dependencies beyond Chart.js — performance by discipline, not by abstraction.
AI · Computer Vision · Generative AI | 2025 – Present
Contributing to applied computer vision research with a focus on generative AI and on-device ML systems. Work spans experimental pipeline design, model evaluation, and translating research findings into deployable, production-ready architectures. Responsible for both original architecture design and iterative critique-and-redesign cycles.
- Designed and shipped a 14-stage production face swap pipeline integrating DreamID, ArcFace, BiSeNet, and CodeFormer
- Executed a rigorous architectural critique and v2 redesign, improving photorealism and identity fidelity under diverse lighting conditions
- Developed a full training suite with modular component separation for reproducible experimentation and isolated ablation testing
- Benchmarked model performance across cross-attention temporal conditioning configurations and quantised variants
Full Stack · ML Engineering · Document Intelligence | 2025
Built an end-to-end compliance automation system leveraging NLP and structured data extraction to streamline regulatory review workflows. Reduced manual document processing overhead for compliance teams through an ML-backed classification and rule-matching pipeline, with a Flutter-based client interface for workflow management.
- Architected a FastAPI backend for document ingestion, parsing, and structured information extraction at scale
- Integrated ML-based document classification to automate compliance rule matching against regulatory corpora
- Built a Flutter client interface enabling compliance officers to review, flag, and approve automated decisions
- Reduced manual review overhead through end-to-end pipeline orchestration from document ingest to decision output
| Recognition | Details |
|---|---|
| 🏆 Hackathon — CrashGuard | Recognised for engineering depth and real-world impact in road safety AI at national hackathon |
| 📊 Competitive Programming | Active Codeforces participant · Problems solved spanning greedy, DP, combinatorics, and graph theory |
| 🔬 Research Engineering | Sole architect of industry-grade 14-stage generative AI pipeline with critique-and-redesign cycle |
| 📚 Academic Breadth | Advanced coursework: DBMS · OS · DAA · Adaptive Signal Processing · Linear Algebra · DWDM |
| 🎯 Investment Banking Pipeline | Candidate for Bank of America Global Investment Banking Summer Analyst (Mumbai, 2027) |
| 🤝 Open Source | Flutter open-source contributor · Full CrashGuard architecture publicly published |
| 🧠 ML Systems | End-to-end ML product ownership: data → model → edge deployment → user-facing application |
-6D28D9?style=flat-square&logo=hackerrank&logoColor=white){kind=icon}
current_focus:
learning:
- "Advanced System Design — Distributed Systems, CAP Theorem, Consensus Protocols"
- "Quantitative Finance & Investment Banking Fundamentals"
- "Transformer Architectures & Vision-Language Models (ViT, CLIP, LLaVA)"
- "Competitive Programming — Codeforces Div. 2 to Div. 1 ladder"
building:
- "Production ML Pipelines — on-device inference + cloud hybrid architecture"
- "Flutter AI-powered cross-platform mobile applications"
- "Personal Engineering Portfolio — glassmorphism, neural canvas, live metrics"
- "Open Source Flutter developer tooling contributions"
exploring:
- "MLOps — model versioning, drift detection, ML CI/CD pipelines"
- "LLM Applications — RAG, function-calling, agentic multi-step workflows"
- "Financial Engineering & data-driven investment analysis frameworks"
- "WebAssembly for high-performance browser-side computation"
open_to:
- "AI / ML Research Internships (India · Remote)"
- "Software Engineering Internships (Full Stack · Mobile · ML)"
- "Open Source Contributions — Flutter, ML tooling, Python libraries"
- "Investment Banking Summer Analyst 2027 (Mumbai)"
Always open to collaborating on meaningful engineering problems. If you're building something that matters — let's talk.
"The code that ships is always better than the architecture that never leaves the whiteboard."