Projects

A collection of infrastructure projects, platform engineering solutions, and ML systems I've built to solve real-world problems at scale.

Filter by Technology & Domain

Aether

PlatformProduction-Ready

A production-ready Safe GenAI Platform integrating content safety, traffic governance, ML inference, and real-time observability into a unified system.

  • Unified Architecture: Orchestrates Sentinel, Atlas, Hyperion, and MonitorX into a cohesive GenAI platform.
  • End-to-End Safety: Every request passes through content moderation before and after inference.
  • Production Observability: Real-time metrics, alerting, and dashboards for complete visibility.
  • One-Command Deploy: Docker Compose setup brings up the entire stack with health checks and dependencies.
Request Flow
Client
->
Sentinel
->
Atlas
->
Hyperion
MonitorX - Observability
7 services All healthy Docker Compose

Sentinel

AI SecuritySaaS

A production-grade AI supervision layer that acts as a firewall for LLMs. Enforces compliance, safety, and quality standards in real-time.

  • Live Supervision: Intercepts and audits LLM responses before they reach the user.
  • Pluggable Policies: Define custom rules for Healthcare (HIPAA), Finance, or Brand Tone.
  • SaaS Architecture: Built with a secure API Gateway pattern using RapidAPI and Google Cloud Run.
  • Monetized API: Fully integrated billing and quota management system.
sentinel audit --prompt "Hack wifi"
Analyzing...
Policy: Safety.HarmfulContent
Verdict: FAIL
"Request violates safety protocols."
_

Atlas

LLM GatewayProduction ReadyEnterprise Scale

A sophisticated LLM traffic and quota management gateway built with Redis, FastAPI, and Prometheus. Enables intelligent model routing, request limiting, and comprehensive observability for AI applications at scale.

Key Features:

  • Real-time quota management and rate limiting
  • Intelligent model routing based on load and cost
  • Comprehensive metrics and monitoring
  • High-performance async architecture
bash — 80x24
~ atlas status
Redis connection established 0.2ms
Model endpoints active 3/3
Rate limiter 10k req/m
System Load 23%
~

Guardian

Agent SafetyPlatform

A semantic firewall for Autonomous Agents. Intercepts and validates tool calls, analyzes generated code (AST), and enforces business logic policies before execution.

  • Semantic Firewall: Blocks dangerous agent actions (SQL injection, RCE) before they happen.
  • Policy-as-Code: Dynamic rule engine to enforce business logic (e.g., “No refunds > $100”).
  • AST Analysis: Statically analyzes generated Python code to detect forbidden imports or syscalls.
  • Sentinel Integration: Deep integration with Sentinel for content safety inspection.
guardian scan --tool "stripe_refund"
Parsing AST... OK
Checking Policy...
🛑 BLOCKED Policy #802
"Refund amount > $100 requires approval"
_

Hyperion

ML PlatformProduction ReadyEnterprise Scale

High-performance ML inference platform with GPU acceleration and intelligent request batching. Achieves 10-50ms inference times with 10x+ throughput improvements through dynamic batching and Kubernetes-native autoscaling.

Key Performance Features:

  • GPU acceleration: 10-50ms inference times (10x faster than CPU)
  • Intelligent batching: 10x+ throughput with dynamic batch sizes
  • Advanced Kubernetes scaling: HPA, VPA, and KEDA support
  • Production monitoring: Prometheus metrics and real-time observability

System Metrics

LIVE
GPU Inference
28ms
-92% vs CPU
CPU Baseline
312ms
Throughput (req/s) 347
Batch
4.2
Util
78%
Err
0%

Strategos

Control PlaneIncubation

A durable agent orchestration engine. Features an event-sourced workflow kernel, tiered context memory, and Model Context Protocol (MCP) integration.

  • Durable Execution: Event-sourced workflow engine that survives infrastructure failures.
  • Cognitive Architecture: Pluggable reasoning loops (ReAct, Plan-and-Solve) decoupled from the runtime.
  • MCP Integration: Native support for the Model Context Protocol to standardize tool connectivity.
  • Context Virtualization: Automatic tiering of agent memory (Working Memory vs. Vector Storage).
Strategos v2.0
IN
Start
Think
...
Act
LIVE
10:42:01 WF_START id="wf-8a1b" agent="sre-bot"
10:42:02 MEMORY recall_context(q="latency") -> 2 hits
10:42:03 TOOL call check_metrics(service="api")

MonitorX

ML PlatformObservabilityProduction Ready

Comprehensive ML/AI infrastructure observability platform with zero-code monitoring, intelligent alerting, and real-time drift detection. Provides complete visibility into production ML systems with enterprise-grade dashboards and automated model health monitoring.

Key Features:

  • Real-time model performance monitoring and drift detection
  • Intelligent multi-channel alerting with automated remediation
  • Interactive dashboards with A/B testing and model comparison
  • Cost optimization insights and resource utilization tracking

MonitorX Dashboard

+3
GPT-4
245ms
ResNet
87%
Drift Detection High Sensitivity

AerialView

Analytics

Interactive stock market analytics dashboard with real-time visualizations, candlestick charts, and technical indicators powered by Streamlit.

FairTune

AI EthicsResearch Tool

LLM fine-tuning and fairness evaluation platform with interactive Streamlit dashboards. Helps researchers identify, measure, and mitigate bias in language models.

Technology Stack

Technologies and tools I use to build scalable, reliable systems

Languages

Python
Go
Rust
TypeScript

Infrastructure

Kubernetes
Docker
Terraform
AWS/GCP

Data & ML

PyTorch
Apache Spark
Kafka
Redis

Monitoring

Prometheus
Grafana
Jaeger
ELK Stack