Research Vision
Corporate technology services website with modular product and services storytelling.
Available for full-time roles — Bengaluru & Remote
Full Stack
Developer
+ Applied AI
Full Stack Engineer specializing in Adversarial AI Security and production-grade LLM architectures.
Ranked Top 8% Globally in the Lakera Gandalf Security Challenge.
0th
Percentile · Global AI
Security Challenge
Security Challenge
0.0%
F1-Score on legal
AI benchmark
AI benchmark
0+
Adversarial techniques
documented & tested
documented & tested
0+
Production apps
shipped end-to-end
shipped end-to-end
01 — Award-Winning Research
Published Research
🏆 Best Paper Presentation — NCRIE-2025
AI-Powered Legal Document Analysis with an Integrated Trust & Safety Framework
Fine-tuned a BERT transformer on 510 CUAD legal contracts to automate clause extraction and risk classification. Integrated a Trust & Safety module with automated PII redaction, live fairness auditing, and privacy analysis — making AI-driven legal review accurate, transparent, and responsible. Full-stack application, not just a paper.
Model Performance
0.0%
F1-Score
0.0%
Accuracy
0
Contracts
CUAD
Benchmark
02 — Work Experience
Where I've Shipped
ASPL Tech Solutions Pvt. Ltd.
Full Stack Developer · Product Delivery Focus
Oct 2025 – March 2026 · Project Tenure
Oct 2025 – March 2026
Bengaluru, India
Executed the full-cycle delivery of 8+ production applications within a high-intensity delivery sprint, managing the complete roadmap from architectural design to cloud deployment.
Built a full-stack HRMS platform digitizing onboarding for 4 departments, reducing administrative processing time by 35%.
Implemented secure RBAC with JWT/bcrypt and optimized RESTful APIs for real-time Recharts dashboards, improving client satisfaction scores by 15%.
Configured and managed 3+ live production sites on Linux VPS using Nginx, handling full SDLC orchestration including SSL/TLS and domain management.
03 — Selected Projects
Things I've Built
AI-Powered Legal Document Analysis
Lead Full Stack AI Engineer · Enterprise Contract Analytics & Trust & Safety Suite
Designed and developed a production-grade, multi-format legal contract analytics platform that automates clause extraction and risk analysis from scanned PDFs, images, and text documents with a robust, integrated Trust & Safety and Algorithmic Fairness compliance pipeline.
The Problem
Manual contract review is slow. However, scanned contracts have watermarks and stamps that corrupt OCR. Moreover, sensitive PII cannot be exposed to external APIs, and neural classifiers exhibit unchecked demographic performance biases.
The Solution
Built an intelligent layout-aware processor programmatically bypassing watermarks, a fine-tuned Legal-BERT classifier, a local spaCy PII redactor generating Privacy Scores, and an auditing system evaluating F1 parity across classes.
Key Engineering Decision
Fine-tuned BERT on the CUAD benchmark instead of relying on prompt engineering alone — because 84.9% F1 is a number you can defend in a paper and in production. Measurable over impressive.
Result
84.9% F1, 84.7% accuracy on CUAD (510 legal contracts). Research paper presented at NCRIE-2025, KSIT — Best Paper Presentation Award.
92.0%
Macro F1-Score
+32%
OCR Accuracy Boost
10+
Clause Categories
< 10%
Bias Disparity Gap
Computer Vision Preprocessing & Neural Classification
Scanned agreements present noisy artifacts. We run page-by-page coordinates pruning using OpenCV and fitz to isolate body boundaries, stripping stamps and page metadata before running adaptive binarization to boost OCR text recovery.
1. OpenCV Core
Boundary Grooming
Thresholding & Noise Removal
→
2. Legal-BERT
PyTorch Fine-Tuning
CUAD Dataset · 6 Epochs · fp16
→
3. Active Learning
HITL MongoDB Collector
User-corrected Annotations Pipeline
Adaptive Document CV
- Programmatically clips top and bottom margin zones of pages, isolating body content and removing e-stamp logos, signatures, and page numbering clutter.
- Applies OpenCV Gaussian smoothing, adaptive binarization, and horizontal/vertical morphological filters to make text fully high-contrast for Tesseract.
Hugging Face Classifier
- Fine-tuned
nlpaueb/legal-bert-base-uncasedusing Hugging Face and PyTorch on CUAD (Contract Understanding Atticus Dataset) text chunks. - Classifies 10+ clauses with logit confidence scores, returning predictive probability distribution margins.
Trust & Safety Pipeline & Fairness Auditor
Privacy is critical. The local NLP redactor removes names, organizations, SSNs, and emails on local servers. A dedicated compliance auditor checks prediction accuracy gaps across document demographics to prevent algorithmic bias.
Dual-Engine PII Redactor
- spaCy + RegEx: Deploys a large neural spaCy Named Entity Recognition model for spatial nouns coupled with optimized regular expressions for numbers.
- Privacy Score: Generates a mathematical privacy risk metric based on entity density, blocking storage of unredacted files.
Fairness Auditing Suite
- Parity Tracking: Computes demographic parity and F1 performance disparities between categories using Scikit-Learn to ensure equitable model performance.
- Bias Warning Flags: Triggers warning indicators if performance gaps between classes exceed a strict 10% threshold.
Stack & SkillsPyTorchPythonOpenCVLegal-BERTspaCy NLPScikit-LearnReact.jsNode.jsMongoDB
PromptRoute
Lead AI/ML & Frontend Engineer · Offline-First Hybrid AI Prompt Router
View RepositoryChrome Extension MV3
Designed and built a dynamic in-browser prompt routing engine that replaces generic, rigid instruction templates with targeted cognitive routing. By running a quantized local neural network, it automatically classifies user intent and wraps inputs in the optimal academic reasoning paradigm (out of 17 supported techniques) in under 15ms.
Monolithic Problem
Most AI assistants rigidly wrap every query in generic templates (e.g. Role -> Task -> Constraint). This wastes expensive tokens on simple chats and lacks targeted cognitive reasoning for complex logic.
Dynamic Solution
PromptRoute intercepts inputs in real-time and runs a local neural network to classify intent, wrapping queries in the precise academic technique (out of 17 paradigms) matching the specific use case.
The Sandbox Constraint
Chrome Manifest V3 service workers cannot execute heavy WebAssembly/WebGPU compilations directly. Bypassed this by mounting ONNX Runtime Web in a persistent Chrome Offscreen Document.
The Hybrid Architecture
Local INT8 classifier executes locally in <15ms. Low-confidence queries (<55%) automatically route to a high-fidelity Groq Llama 3.3 70B cloud API, securing 80% token cost savings.
< 15ms
Local Latency
14.8MB
Quantized Size
85.6%
Eval Accuracy
7,477
Dataset Samples
~80%
Token Savings
Python & Google Colab ML Pipeline
Engineered an end-to-end machine learning pipeline in Python to fine-tune distilbert-base-uncased for local browser-side inference. The dataset pipeline balances diverse instruction-following and coding prompts with engineered casual colloquialisms to ensure robust real-world intent matching.
1. Data Engine
7,477 Prompt Samples
Dolly, Alpaca, Glaive + Custom Synthetics
→
2. Colab PyTorch
Tesla T4 GPU Fine-Tuning
AdamW · fp16 · 6 Epochs · 85.6% Acc
→
3. ONNX Quantization
14.8MB INT8 Compile
Offline WASM/WebGPU Inference
Data Engine & Preprocessing
- Dataset Blending: Merged key programming prompts from Glaive, task instruction-following from Alpaca, and factual QA from Databricks Dolly-15k.
- Synthetic Augmentation: Programmed over 1,000 synthetic samples targeting local linguistic idioms, typographical errors, and casual Indian English phrasing (e.g., "kindly draft...").
Model Architecture & Quantization
- Training Configuration: Leveraged PyTorch and Hugging Face to fine-tune DistilBERT over 6 epochs, utilizing AdamW optimizer, warmup ratio of 0.1, and mixed-precision (
fp16). - INT8 ONNX Compressing: Compiled PyTorch weights to an open-format ONNX graph, quantizing weights to 8-bit integers to compress the model from ~270MB down to 14.8MB.
Chrome MV3 Extension & Fallback Runner
Bypasses Manifest V3 service-worker execution environments by running in-browser inference using ONNX Runtime Web in a sandboxed, hidden Offscreen Document. This architecture loads the WASM/WebGPU models eagerly upon extension boot to prevent service-worker eviction or freeze issues.
Offscreen Sandbox & Messaging
- WASM & WebGPU SIMD: Leverages hardware WebGPU acceleration in the sandbox offscreen page to perform local classification in under 15ms.
- IPC Native Bridging: Communicates between host DOM panels, Manifest V3 background service worker threads, and the sandbox using low-latency Chrome native messages.
Hybrid Fallback & UI
- Secure Groq Runner: Intercepts low-confidence intent classifications (<55%) and routes them to a cloud Groq runner (Llama 3.3 70B) to guarantee high-fidelity reasoning.
- Premium Glassmorphic UX: Injects a translucent, responsive control overlay with backdrop filters, manual overrides, and live token estimators.
Stack & SkillsPythonLLM Fine-TuningPyTorchHugging Face TransformersJavaScript ES6+ONNX Runtime WebWebGPUChrome MV3Jest
AI Security Lab
Prompt Injection & Defensive Engineering Lab
A comprehensive adversarial testing environment where I neutralized all 7 levels of the Lakera Gandalf challenge. I achieved a Top 8% Global Rankingby developing a "Defense-in-Depth" framework.
The Vulnerability
LLMs are susceptible to sophisticated prompt injections that easily bypass standard input sanitization and moderation layers.
The Defense
Documented 15+ adversarial techniques (Multilingual Obfuscation, Data Fragmentation) and engineered intent-based guardrails to expose and patch model blindspots.
Stack & SkillsLLM GuardrailsAdversarial TestingRed TeamingPrompt Engineering
Blood Donor Management System
Full Stack Developer · Healthcare Information & Analytics Platform
Admin dashboard for donor/recipient CRUD operations, automated blood group compatibility matching, donation lifecycle tracking, and analytics-ready medical database management.
The Problem
Managing donor profiles, matching blood types during emergencies, and updating donation statuses across regional healthcare collection sites was highly manual, sluggish, and prone to tracking errors.
What I Built
A fast, full-stack medical registry system with JWT authentication, donor/recipient directories, direct blood type matching algorithms, donation workflow monitors, and dynamic analytics dashboards.
Key Engineering Decision
Selected MongoDB for its flexible schema (easily absorbing varied donor medical flags) paired with Python-Flask APIs to encapsulate matching logic, ensuring rapid queries and extreme service maintainability.
The Result
Delivered a high-performance system for swift emergency matching, highly indexable profiles, and real-time donation tracking with clean analytics reports for clinical operations.
Stack & SkillsReactFlask (Python)MongoDBREST APIsJWT SecurityViteBootstrap
06 — Skills
Technical Competence
Frontend
Backend
Data & AI
Advanced Prompting & Security
DevOps & Tooling
React.js
Practical Implementation:Core technology used for ASPL client portals, custom dashboard widgets, and responsive B2B platform layouts.
07 — About
Who I Am
I'm a Full Stack Developer based in Bengaluru. My work sits at the intersection of modern web engineering and applied AI — I design schemas, integrate models, configure servers, and ship UIs that real people use.
I served as a Full Stack Developer at ASPL Tech Solutions, where I took ownership of production repositories and client deliveries from day one.
My research on AI-powered legal document analysis won Best Paper at NCRIE-2025. That work wasn't an academic exercise — it's a deployed, full-stack application backed by a fine-tuned BERT model trained on 510 real contracts.
I hold an MCA from RV Institute of Technology and Management (CGPA: 8.2). I'm actively looking for a role where I can keep building things that matter.
Award
🏆
Best Paper Presentation
NCRIE-2025 · KSIT, Bengaluru
National Conference on Recent Innovations in Engineering
View Certificate →National Conference on Recent Innovations in Engineering
Certification
Certified MERN Stack Developer
Dyashin Technosoft
