Initializing Experience...
Portfolio / 2026
Gaurav
Kumar
Nayak
Full Stack Developer & AI Engineer
B.Tech CSE · C.V. Raman Global University · 2024–2028
Scroll to explore
I turn
magic into
logic.
AI isn't magic — it's architecture, iteration, and obsession. I just make it look effortless.
I engineer
what others
imagine.
Where curiosity meets code — designing intelligent systems that don't just work, but make people wonder how.
See what I've built
Who I Am
3rd-Year CS Student.
AI Systems & Full-Stack Engineer.
Building production-grade systems and agentic workflows.
Amazon Prize
Nova AI Hackathon '26
Top 800 / 31k+
Meta × OpenEnv
12 Built
6 Deployed, 6 Prototypes
I'm Gaurav — a computer science student at C.V. Raman Global University specializing in AI engineering and full-stack development. I focus on building reliable software: multi-agent LLM pipelines, custom RAG architectures, and distributed backends. My work has been awarded at Amazon, Meta, and various IIT hackathons. I'm actively seeking AI engineering internships and remote developer roles where I can build real systems from day one.
Current Research Focus
Active research vs past skills — May 2026
Mechanistic Interpretability
Understanding what transformer attention heads actually encode. Using CircuitScope + GemmaScope + TransformerLens to run ablation studies on coding-related circuits. Goal: explain why fine-tuned models improve on specific benchmarks.
Evaluation-First LLM Development
Building evaluation suites before writing a single model prompt. Combining RAGAS (faithfulness, relevancy), DeepEval (hallucination, toxicity), and custom HumanEval variants for code-specific benchmarks.
Stateful Multi-Agent Systems
Moving beyond single-turn agents to persistent, stateful workflows using LangGraph. Focus on self-correction, contradiction detection, and graceful failure modes for production deployments.
GPU Optimization & Triton Kernels
Writing custom Triton kernels and fusing layers in SRAM to bypass memory bottlenecks and optimize inference speed on standard GPUs.
Available for Summer 2026 AI internships and remote junior roles
Currently open to AI engineering internships, remote developer roles, research collaborations, and high-impact full-stack opportunities.
Impact in Numbers
Rigorous engineering benchmarks and competitive milestones validating system quality
GPU KERNEL FUSION SPEEDUP
Achieved 8.2x speedup on RMSNorm compared to eager PyTorch implementation by fusing GPU memory accesses in SRAM.
QUALEDGE NPU LATENCY
Delivers 0.55ms latency and 41x speedup executing 100% NPU-native on Hexagon HTP via AIMET 8-stage PTQ quantization.
CODESAGE VLLM THROUGHPUT
High-throughput serving of QLoRA 4-bit fine-tuned models via PagedAttention with custom MCP server integration.
NEURALVAULT RAG FAITHFULNESS
Achieved premium retrieval fidelity via hybrid BM25 + pgvector Reciprocal Rank Fusion, evaluated via RAGAS framework.
RAILMIND AUTONOMOUS DISPATCH
Coordinates a 6-agent LangGraph pipeline with XGBoost RAC conflict prediction and PPO RL dynamic rescheduling.
HYPERFLOW DEMAND FORECAST
Achieved +24.28% WMAPE forecast lift on stockout-censored inventory using Tobit MLE loss with Swiggy MCP agents.
AGENTSENTRY CACHE SAVINGS
Slashes LLM token costs by 50% using suffix-delta caching and guards execution via AST exploit shielding.
PRODUCTION SYSTEMS
Engineered stateful multi-agent systems, edge AI quantizers, GPU kernels, and distributed commerce engines.
Amazon Nova AI Hackathon
Won exclusive Bonus Blog Post Prize ($200 AWS Credits) — the only team to receive this award. Built multi-agent AI research automation.
Meta × OpenEnv Hackathon
Ranked in the Top 800 globally out of 31,000+ teams in a flagship reinforcement learning challenge judged by Meta AI researchers.
Professional Credentials
15+ High-impact certifications from Stanford, IBM (GenAI Specialist), Meta (Full Stack), and IITs in AI/ML & System Design.
GitHub Contributions
Every day commit heatmap tracking active AI orchestration pipelines and MLOps integrations
Technical Proficiency
Recruiter-focused capability indexes showing depth and hands-on production expertise
Proficient
Built 2+ working systems, can explain architecture and tradeoffs live, and can debug under pressure.
Working Knowledge
Successfully used in at least one project, understand core concepts, but refer to documentation for edge cases.
Learning / Exploring
Actively studying, running experiments and tutorials, but have not yet deployed independently.
Developer Growth Arc
Foundation & Problem Solving
Mastered Python, C++, and core data structures. Solved 100+ LeetCode problems and began building automated scripts and CLI workflows.
Full Stack Web & Early ML
Transitioned to professional web architectures and predictive machine learning models. Learned state management, API designs, and dataset vectorization.
Architecture, Caching & Scalability
Focused on high-performance distributed systems, low-latency microservices, relational range partitioning, and advanced indexing strategies.
Agentic AI & Mechanistic Interpretability
Building production LLM systems. GPU optimization via Triton, mechanistic interpretability causal patching, multi-agent LangGraph flows, and serverless MLOps infrastructure.
AI Engineering Internship at FlyRank
Building automated SEO and content optimization pipelines using multi-agent LLM systems, structuring schema markup for AI-powered search engines, and configuring llms.txt standard files to optimize visibility.
What I Build
Systems I've shipped and the problems they solved
Amazon Prize WinnerMulti-Agent LLM Pipelines
I design and ship autonomous agent systems using LangGraph and CrewAI. ResearchPilot orchestrates 5 specialized nodes with persistent state, self-correction, and contradiction detection — delivering cited research reports in under 5 minutes.
8.2x SpeedupGPU Kernel & Systems Engineering
I write custom GPU kernels using OpenAI Triton and PyTorch JIT to bypass memory-bandwidth bottlenecks in LLM inference. TritonForge fuses memory-bound activations directly within SRAM, delivering CUDA-equivalent throughput with 99.2% HBM memory savings.
8-System ArchitectureProduction Full-Stack Systems
I build distributed platforms with React 19, Next.js, and FastAPI. CineNexus handles global watch parties, real-time WebSocket sync, and AI-driven personalization across an 8-system microservices architecture.
87% RAG FaithfulnessRAG Architectures & Vector Search
I implement hybrid retrieval systems combining dense pgvector embeddings with sparse BM25. LexRAG achieved 87% RAGAS faithfulness vs 61% for naive RAG, using GraphRAG over a Neo4j knowledge graph.
Tree SHAP ExplainabilityMLOps & Production FinTech Observability
I build production-grade FinTech observability and intelligence platforms. Artha AI features supervised ensemble fraud classification, real-time Tree SHAP attributions, SEBI compliance circular RAG reranked with cosine similarity, and real-time MLOps data drift auditing.
GemmaScope SAEs · KL PatchingNeuroScope: Agentic Interpretability
I build mechanistic interpretability platforms. NeuroScope captures residual streams, attention patterns, and SAE features at every step of a multi-turn agent, enabling cross-step causal activation patching and representation steering.
AST Exploit Shield · 50% Cache SavingsAgent Security & Prompt Firewalls
I design secure agent runtimes and optimization middleware. AgentSentry guards remote code execution via recursive AST parsing, intercepts subshell breakouts, and slashes token costs by 50% using suffix-delta caching.
+7.1pp HumanEval · 12,500 tok/sQLoRA Fine-Tuning & LLM Serving
I fine-tune domain-specific code LLMs using QLoRA DoRA NF4 4-bit quantization on 52K instruction pairs (+7.1pp HumanEval). CodeSage serves models via vLLM PagedAttention at 12,500 tok/s on H100 with an MCP server exposing complete, review, and test-generation tools.
0.55ms NPU Latency · 41× SpeedupSnapdragon Edge AI & AIMET Quantization
I optimize and deploy NPU-native ML pipelines to edge silicon using Qualcomm AIMET 8-stage PTQ and AI Hub compilation. QualEdge executes 100% NPU-native on Hexagon HTP with 0 CPU fallbacks, delivering 0.55ms latency and 194.62× model compression.
6-Agent LangGraph · ECDSA AuditAutonomous Dispatch & RL Control Systems
I architect mission-critical AI control platforms for large-scale physical networks. RailMind coordinates a 6-agent LangGraph dispatch pipeline for Indian Railways with XGBoost RAC conflict prediction, BFS delay cascade propagation, PPO RL dynamic rescheduling, and an ECDSA cryptographic audit chain.
+24.28% WMAPE · Swiggy MCP AgentHyperlocal Commerce Intelligence & Production ML
I solve documented production ML gaps in supply chain and e-commerce. HyperFlow handles stockout-censored demand via Tobit MLE (+24.28% WMAPE lift), velocity-normalized RF gates suppressing ETA display jitter by 81.4%, and Gemini 2.0 Flash ReAct agents with live Swiggy MCP APIs.
Global Recognition & Ecosystem
Recognized At
Competitions, certifications, and programs hosted by India's leading institutions and global tech companies
Featured Projects
Selected systems engineering and AI research projects benchmarked on production performance metrics.
ResearchPilot
Amazon Prize Winner · Multi-Agent LLM Orchestrator
The Problem
“Manual research synthesis takes 4–6 hours of tab-switching and cross-referencing across dozens of sources.”
What I Built
“5-node LangGraph pipeline with Search, Read, Reflect, Cross-Check, and Write agents. Self-corrects via reflection loops. Cites every claim.”
Impact
Primary Tech Stack
RailMind
Autonomous Dispatch Intelligence · Indian Railways
The Problem
“Indian Railways dispatch runs on fragmented manual coordination — conflict resolution, delay cascade tracking, and rescheduling are siloed with no unified AI decision layer.”
What I Built
“6-agent LangGraph orchestration pipeline with XGBoost RAC (Route Acquisition Conflict) prediction, BFS cascade propagation for delay chain simulation, PPO reinforcement learning for dynamic rescheduling, and an ECDSA cryptographic audit ledger for tamper-proof decision logging.”
Impact
Primary Tech Stack
HyperFlow
Hyperlocal Commerce Intelligence · 4 ML Problems Solved
The Problem
“Swiggy Bytes and Zomato Engineering document four unsolved production ML problems: censored demand from stockouts, ETA display jitter from GPS noise, cancelled order resale arbitrage, and refund loop fraud from cloud-kitchen geo-collision. No existing platform solves all four.”
What I Built
“Production ML platform solving each gap from first principles: Heteroscedastic Tobit MLE + LightGBM Quantile for censored demand (+24.28% WMAPE), velocity-normalized RF Classifier gate for ETA jitter (81.4% suppressed), Thermal SQI + Sybil proximity guard for arbitrage (100% blocked), tenure-gated bypass + semantic plausibility engine for fraud (0% false positives). Gemini 2.0 Flash ReAct agent with live Swiggy MCP APIs. Atomic dual-mode locking (Redis Redlock + PostgreSQL FOR UPDATE NOWAIT) at 1,598 req/sec.”
Impact
Primary Tech Stack
CodeSageZ
Graph-Augmented RAG · Tree-sitter AST · QLoRA 4-bit
The Problem
“General-purpose code LLMs underperform on domain-specific repository completion and review tasks, and naive vector search misses multi-hop structural call dependencies.”
What I Built
“Built a Retrieval-Augmented Generation (RAG) engine that expands vector seeds via 1-hop BFS over a Tree-sitter AST call graph and re-ranks with a hybrid score (0.6×cosine + 0.4×graph proximity). Fine-tuned Qwen2.5-Coder-1.5B via 4-bit QLoRA (Unsloth) on 8K CommitPack bug-fix commits.”
Impact
Primary Tech Stack
QualEdge
Qualcomm Snapdragon Edge AI · 0.55ms NPU · 41× Speedup
The Problem
“No production-grade open platform demonstrates the full Qualcomm ML stack end-to-end — from AIMET quantization through AI Hub compile and real Snapdragon NPU profiling — with reproducible, sourced benchmarks.”
What I Built
“Full 8-stage AIMET PTQ & AI Hub compilation pipeline (BN fold, CLE, ReLU surgery, AdaRound INT8, QNN compile job j5w110q4g, HTP profile job jgdzzyo65). Achieves 100% Hexagon HTP-native execution with 0 CPU fallback operators. Hybrid on-device/cloud LLM router (ModernBERT + LogReg) with 1.85ms p95 router latency and 100% local HTP cloud avoidance.”
Impact
Primary Tech Stack
CineNexus
8-System Distributed Watch Parties Portal
The Problem
“Friends watching movies together across time zones have no synchronized, social streaming experience with personalization.”
What I Built
“8-service distributed platform: Next.js frontend, FastAPI backend, WebSocket sync engine, Redis pub/sub, AI recommendation layer, and PostgreSQL.”
Impact
Primary Tech Stack
TritonForge
GPU Kernel Engineering · OpenAI Triton (Up to 8.2× Faster)
The Problem
“Memory-bound LLM bottlenecks (like SwiGLU activations and RMSNorm) limit throughput and waste high-bandwidth memory (HBM) on modern GPUs.”
What I Built
“A suite of fused CUDA-equivalent GPU kernels written in OpenAI Triton including SwiGLU, RMSNorm, and FlashAttention-2. Fuses memory-bound operations directly within GPU SRAM, breaking the memory wall to deliver up to 8.2× speedup over PyTorch eager mode.”
Impact
Primary Tech Stack
Artha AI
Tree SHAP Explainability · RegGuard RAG Platform
The Problem
“Financial ensemble models predict fraud with high accuracy, but compliance auditors cannot act on them without explainable features and regulatory alignment.”
What I Built
“Voting ensemble (XGBoost + LightGBM + Random Forest) + real-time Tree SHAP values + pgvector compliance document matching RAG and LangChain SQL agents.”
Impact
Primary Tech Stack
Also Notable Work
Other shipped systems, ML pipelines, and competitive hackathon builds
6-agent LangGraph AI dispatch intelligence for Indian Railways. Features XGBoost RAC conflict prediction, BFS delay cascade propagation, PPO dynamic rescheduling, and an ECDSA cryptographic audit ledger.
Hyperlocal commerce intelligence platform solving 4 documented Swiggy/Zomato ML gaps: Tobit MLE demand (+24.28% WMAPE), RF ETA jitter gate (81.4% suppressed), SQI arbitrage shield, and Gemini 2.0 Flash ReAct agent with live Swiggy MCP.
QLoRA fine-tuned Llama 3.3 8B model on 52K code pairs (+7.1pp HumanEval). Served via vLLM PagedAttention at 12,500 tok/s on H100 with a 4-tool MCP server for Claude Desktop and Cursor.
Full Qualcomm AIMET PTQ pipeline targeting Snapdragon X Elite NPU with 0.55ms latency (41× speedup). Features hybrid on-device/cloud LLM router classifier at 93.3% accuracy and real-time telemetry.
Memory-bound LLM bottleneck solver using OpenAI Triton. Fuses memory-bound activation layers and RMSNorm directly within GPU SRAM, eliminating HBM read/write overhead to deliver CUDA-equivalent performance.
FinTech audit platform featuring ensemble transaction fraud matching, real-time Tree SHAP explainability, online SEBI/NPCI circular RAG, and LangChain SQL statement auditing.
Mechanistic interpretability platform mapping LLM agent trajectories. Decomposes residual streams using 16k-width GemmaScope Sparse Autoencoders and runs causal activation patching to isolate reasoning failures.
A legal AI research assistant utilizing dense pgvector embeddings with sparse BM25 and Neo4j GraphRAG. Achieves 87% faithfulness, evaluated using the RAGAS framework across five custom, modular RAG pipelines.
A stateful multi-agent customer support ticket triage and routing platform built for Meta × OpenEnv. Spawns specialized CrewAI solvers in a LangGraph pipeline to route and solve tickets dynamically.
A distributed movie recommendation and discovery system handling watch parties, real-time playback sync, and AI-driven personalization across an 8-system microservices architecture.
AI-driven women's safety platform providing real-time threat detection via Whisper + BERT ambient voice monitoring, one-tap emergency SOS calls, location tracking, and visual safety heatmaps.
Multi-agent financial operating system running CrewAI solvers to execute real-time spend intelligence, 13-week cash flow forecasts, month-end reconcile loops, and transaction auditing.
AI-adaptive onboarding engine that parses resumes + JDs via Gemini, maps skills gaps to the O*NET taxonomy, and dynamically generates prerequisite-ordered pathways via Kahn's topological sort.
Bilingual ambient medical AI scribe for Indian clinics. Converts Hindi + English doctor-patient consultations into FHIR R4-compliant clinical summaries in under 2.5 seconds using BioBERT NER.
Real-time AI avatar interview system. Integrates custom D-ID avatars, Whisper audio transcribers, MediaPipe facial markers analysis, and XGBoost models for multi-dimensional grading.
ML-driven GNSS spoofing detector utilizing hybrid LSTM Autoencoder + XGBoost models. Processes 18 high-frequency features on TEXBAT signals with 97%+ recall and under 15ms latency.
Security science and prompt caching middleware for autonomous agents. Intercepts AST shell exploit attempts and reduces LLM API costs by 50% using suffix-delta caching.
Hackathons & Achievements
Amazon Nova AI Hackathon
Amazon Web Services · Devpost
Project Shipped
ResearchPilot (Multi-Agent Research Pipeline)
“Manual research synthesis takes 4–6 hours of tab-switching and cross-referencing across dozens of sources.”
“5-node LangGraph pipeline with Search, Read, Reflect, Cross-Check, and Write agents. Self-corrects via reflection loops. Cites every claim.”
Amazon Prize Winner · AWS Blog Post Published · Sub-5min reports
Meta × OpenEnv Global AI Agent Hackathon
Meta AI · HuggingFace · Scaler
Project Shipped
SupportOps-Env (RL environment for AI agents)
“Customer support at scale requires consistent, policy-compliant responses but human agents produce inconsistent quality under volume.”
“Multi-agent support automation: intent classifier, policy retrieval (RAG), response generator, quality checker, and escalation router.”
Top 800 / 31,000+ globally — Meta × HuggingFace Challenge
ArtPark CodeForge Hackathon
IISc Bangalore
Project Shipped
SkillBridge (Gemini topological sort onboarding)
“Corporate onboarding programs offer generic reading tracks rather than customized, skill-gap aligned pathways.”
“SkillBridge corporate onboarding engine: maps JD gap analyses via Gemini and plots learning paths via Kahn's topological sort.”
Winner — ArtPark CodeForge IISc Bangalore
Economic Times AI Hackathon
The Economic Times
Project Shipped
MarketMind (Multi-Agent Market Analyst)
“Analyzing NSE & BSE market opportunities requires synthesis of massive data streams (news, filings, indicators, and holdings) under intense time pressure, which takes hours for retail investors.”
“Multi-agent AI analyst console coordinating 4 specialized agents (Market Data, Technicals, ET News/BSE Filings, and Portfolio-aware) to synthesize stock investment responses in under 30 seconds.”
Semi-Finalist (Problem Statement 6) — Economic Times GenAI Hackathon
Neural Nexus IIT Jammu
IIT Jammu
Project Shipped
FloatChat (ARGO RAG Oceanographic Widget)
“The ARGO program generates terabytes of complex oceanographic data that usually requires specialized tools and expert domain knowledge to interpret, blocking general accessibility.”
“FloatChat oceanographic assistant: semantic FAISS retrieval (all-MiniLM-L6-v2) for top profiles, Groq LLaMA-3 natural language queries, and 6 interactive auto-triggered chart types (SST time series, depth profiles, T-S diagrams).”
Round 1 Finalist — Neural Nexus, IIT Jammu · Ocean data without domain expertise
DevFusion IIT Bombay × IIT Delhi
IIT Bombay · IIT Delhi
Project Shipped
PrepGrid (AI Interview Platform)
“Off-the-shelf mock interview tools offer static question prompts and lack dynamic domain evaluation feedback.”
“PrepGrid mock interview platform: real-time voice-to-text with Whisper, topological skills scoring, Next.js SSR, and FastAPI backend sockets.”
DevFusion Finalist · 200+ mock scenarios handled during active trials
HackMatrix 2.0 IIT Patna
IIT Patna
Project Shipped
VaidyaScribe (FHIR Ambient Scribe)
“Rural Indian clinics require fast, bilingual medical scribing, but manual record entry takes hours and causes clinical reporting delays.”
“Mobile-first VaidyaScribe dictation: transcribes Hindi/English Whisper dialogues, indexes BioBERT NER, and formats FHIR R4 reports.”
Finalist — HackMatrix 2.0 IIT Patna
Devcation Delhi 2026
Devcation Delhi
Project Shipped
ShieldHer (Ambient Safety System)
“Passive ambient tracking is rarely integrated with audio triggers, delaying emergency alarms during sudden voice-only threat conditions.”
“Passive ambient threat monitor processing speech to text via Whisper, classifying distress intent using BERT, and triggering location-linked alerts.”
Finalist — Devcation Delhi 2026
Technex'26 IIT BHU
IIT BHU
Project Shipped
InterviewX (AI Avatar Evaluator)
“Traditional technical screening interviews lack standardized bias-free grading and real-time behavioral insights.”
“Real-time AI avatar interview system with D-ID avatars, Whisper audio transcription, and MediaPipe facial markers.”
Finalist — Technex'26 IIT BHU
IIT Bombay Kaizen × ARIES × NyneOS
IIT Bombay
Project Shipped
GNSS Anti-Spoofing (LSTM Security)
“ML-driven GNSS anti-spoofing pipeline using a hybrid LSTM Autoencoder and XGBoost. Processed 18 high-frequency features on TEXBAT signal datasets.”
Zaggle × COMET'26 IIT Roorkee
IIT Roorkee · Zaggle
Project Shipped
CFO-OS (Multi-Agent OS)
“Multi-agent spend intelligence: autonomous cash flow forecasting, 13-week cash projections, and double-entry transaction auditing.”
Courses & Certifications
Featured Specializations & Professional Programs
Meta Full Stack Developer Specialization (10-course)
“Meta's official 10-course full stack program used internally at Meta engineering teams. Covers React, Django, MySQL, REST APIs, version control workflows, and UX principles. 240+ hours — directly applied to CineNexus and PrepGrid.”
Key Skills
Machine Learning Specialization
“Andrew Ng's flagship 3-course ML program covering supervised learning (linear/logistic regression, SVMs), unsupervised learning (clustering, anomaly detection), recommender systems, and reinforcement learning. The gold standard ML certification — 100+ hours.”
Key Skills
IBM Data Science Professional Certificate
“10-course end-to-end data science program: Python, SQL, data visualization with Matplotlib and Seaborn, ML with Scikit-learn, and model deployment on IBM Cloud. Capstone: IBM HR Attrition Prediction achieving 0.93+ AUC.”
Key Skills
Generative AI Engineering with LLMs (7-course)
“7-course IBM program covering LLM architecture, RAG pipelines, fine-tuning with PEFT/LoRA, prompt engineering, LangChain agent development, and deploying AI applications with IBM Watson. Directly applied to CircuitScope and NeuralVault projects.”
Key Skills
IBM Full Stack Software Developer Specialization (12-course)
“12-course professional certificate covering cloud-native full stack development with React, Node.js, Python/Flask, Django, REST APIs, containers, Kubernetes, microservices, and CI/CD pipelines. Capstone deployed on IBM Cloud.”
Key Skills
Targeted Courses & Skill Certifications
Ultimate AWS Certified Cloud Practitioner CLF-C02
Udemy · 20 Hours
Complete AWS Cloud Foundations prep by Stéphane Maarek covering core AWS services (EC2, S3, Lambda, RDS, IAM), cloud architecture principles, security, pricing models, and the Well-Architected Framework. Applied to NeuralVault and ResearchPilot deployments.
Programming with Generative AI
IIT Guwahati × Coursera · 24 Hours
IIT Guwahati's GenAI program covering generative AI fundamentals, prompt engineering patterns, LLM architecture internals, and building production AI applications with IBM Watson and open-source models.
Advanced Learning Algorithms
Stanford × DeepLearning.AI · 30 Hours
Deep dive into neural networks built from scratch in NumPy and TensorFlow, decision trees, random forests, XGBoost, and practical ML advice on diagnosing bias/variance and improving model performance. Course 2 of the ML Specialization.
Verified Recognition
Verifiable endorsements of architectural execution, AI performance optimizations, and technical design from peer reviews and global hackathons.
“Building the IBM HR Attrition Prediction System was a masterclass in data-driven problem solving. We engineered complex ML models and interactive dashboards that unlocked real-world insights.”
“ResearchPilot won the Bonus Blog Post Prize for its deep technical breakdown of multi-agent LLM orchestration using Bedrock. Gaurav's system is highly detailed, handling state, self-correction, and contradiction detection.”
Writing & Technical Case Studies
Technical deep-dives into LLM engineering, interpretability research, and full-stack architecture.
ResearchPilot: Turning Hours of Research into Minutes with Amazon Nova
Exploring the multi-agent orchestration of Nova Micro, Lite, and Pro for high-speed cross-source research and contradiction detection. Won the Bonus Blog Post Prize at Amazon Nova AI Hackathon.
TritonForge: Fusing SwiGLU and RMSNorm Kernels in OpenAI Triton
A deep dive into custom GPU memory alignment, SRAM tile optimization, and reducing HBM read/write roundtrips for LLM activation layers. Achieved 3.2x average speedup.
NeuralVault: Building a GraphRAG Retrieval Pipeline on Neo4j & pgvector
Architecting hybrid search (dense pgvector + sparse BM25) and connecting multi-hop entity graphs in Neo4j to boost RAG faithfulness to 87%, evaluated via RAGAS.
Artha AI: Explainable Machine Learning & Regulatory RAG in Production FinTech Systems
Designing secure financial transaction pipelines with ensemble fraud voting classifiers, Tree SHAP attributions under 20ms, and pgvector compliance document matching.
Got an idea? Let's build it.
Open to AI engineering internships, remote developer roles, and technical collaborations.
Or reach me directly at gauravnayak711@gmail.com