MarTech Solutions Architect · Customer Data Platforms · Data Engineering · AI Engineering Bengaluru, India
I build open-source tools for the parts of the customer-data stack that break quietly: tags that stop firing, consent that does not actually block, CRM data that drifts, and scripts that tax every page.
Start here · Open source · Technologies · Contact
| Project | What it does |
|---|---|
| PyTagManager | Debug why GTM tags do not fire. Inspects containers, correlates runtime DOM → dataLayer → GTM → network in a real browser, audits consent and script performance, and generates reviewable fixes. PyPI |
| QuillSift | Rust-native, self-hosted scraping engine: SSRF-guarded fetching, deterministic HTML → Markdown/text/JSON, and schema-validated LLM extraction (OpenAI-compatible + Ollama). PyPI |
| PyTagShield | Consent-aware, privacy-minimizing event pipeline with a Rust core and Python API. Validate, pseudonymize, minimize, and route events to GA4, Meta CAPI, PostHog, sGTM, or webhooks. PyPI |
| Squid-CRM | Self-hostable CRM backend, API-compatible with Twenty. Rust + Axum + PostgreSQL + GraphQL/REST, multi-tenant, RBAC, custom objects, webhooks. Alpha. |
Also worth a look: OmniPilot — verification-first mobile automation with evidence-backed results and an MCP server · TinyBridge — a macOS-native Linux VM runtime that boots a real Ubuntu guest on Apple's Virtualization.framework · PyStreamXL — constant-memory Excel streaming with a Rust core.
- Measurement you can verify. Tags, dataLayer contracts, consent, and the gap between "the tag fired" and "the event arrived correctly."
- Customer data that stays usable. Identity, segmentation, reverse ETL, and the pipeline reliability that CDPs depend on.
- Local-first tooling. Diagnostics that run on your machine, with honest confidence labels instead of fabricated certainty.
Solutions Architect with more than a decade of experience across IBM, Wipro, IPG Mediabrands, and others, designing and delivering enterprise-scale customer data, marketing technology, analytics, and AI-enabled solutions. Expertise in Customer Data Platforms (Adobe RT-CDP, Braze, mParticle, Segment CDP), event-driven architectures, cloud data platforms, customer journey orchestration, identity resolution, real-time activation, and data engineering on Google Cloud and Azure.
Outside of MarTech, I build systems-level software in Rust for fun, including a from-scratch AI-native OS kernel (the SHER OS family).
- Background: MBA, Jansons School of Business · B.Com, Mar Ivanios College
- Exploring: agentic AI for MarTech workflows, RAG architectures, and MCP-based agent tooling
Every project links straight to its Use Cases section where one exists — the fastest way to see whether it solves a problem you actually have. Status is stated plainly: if something is early-stage or has known gaps, the README says so.
Not sure which project you need? Several tools share deliberately similar names — the
Py*Guard,Py*Check,Py*Hound, andPy*Meshfamilies. Each of those repos opens with a short "Which project do I need?" table that maps common problems to the right tool.
| Repo | What it does |
|---|---|
| PyTagManager | Debugs GTM implementations: container analysis, semantic DOM graph and tracking recommendations, real-browser runtime diagnosis, site-wide tagging QA, data dictionary, tag execution graph and test candidates, consent audit, and controlled script-impact experiments. |
| PyTagShield | Consent-aware, privacy-minimizing server-side event pipeline with a Rust core: validation, pseudonymization, data minimization, and destination routing (GA4, Meta CAPI, PostHog, sGTM, webhooks). |
| ClusterAudienceKit | Rust-powered customer segmentation: RFM, KMeans/K-Prototypes with automatic K selection, churn/CLV, lookalikes, cohorts, drift detection, streaming segmentation, SQL export to 8 dialects, differential privacy/k-anonymity, and revenue/price intelligence. |
| PyReverseETL | Real reverse ETL with Postgres/MySQL/S3/webhook and Salesforce/HubSpot/Marketo/Squid-CRM connectors, PII masking, validation gates, retry/idempotency, and a queryable lineage graph. |
| PyCustomerHub | Self-hosted CDP and engagement platform: identity resolution with provenance, real-time audiences, governance and consent-aware activation, durable journeys, Postgres + Kafka, and Web/iOS/Android/JS/Python SDKs. |
| Squid-CRM | Self-hostable CRM backend, API-compatible with Twenty. Multi-tenant, RBAC, custom objects, GraphQL/REST, and webhooks. Alpha; the frontend is not vendored. |
These projects are complementary by design — each is useful on its own, but together they cover a customer-data lifecycle end to end:
- PyTagManager — the measurement layer: implements and runtime-verifies the analytics tags that generate the events.
- PyTagShield — the consent and routing layer: minimizes what leaves the browser and controls which destinations receive it.
- PyCustomerHub — the ingestion and engagement layer: resolves identity, evaluates audiences in real time, governs activation, and runs durable journeys across Web, iOS, Android, JS, and Python SDKs.
- ClusterAudienceKit — the intelligence layer: turns raw data into segments with RFM analysis, clustering, and churn/CLV scoring.
- PyReverseETL — the activation layer: pushes segments to CRM and marketing-automation destinations with PII masking and lineage.
- Squid-CRM — the system of record layer: a self-hostable CRM backend for the data the rest of the stack activates into.
The stack reads as one round trip: verify capture → enforce consent → ingest → segment → activate → land in CRM — and each project is built to snap into the others rather than compete with them.
| Repo | What it does |
|---|---|
| QuillSift | Rust-native, self-hosted scraping: SSRF-guarded fetching, MudFish-backed crawling, deterministic HTML → Markdown/text/JSON, schema-validated LLM extraction, Python bindings, and a CLI. |
| MudFish | Rust-native async crawler: bounded budgets, post-resolution SSRF protection, robots.txt and per-host politeness, resumable crawls, and opt-in headless-Chrome rendering; Python bindings via PyO3. |
| Repo | What it does |
|---|---|
| PyAirflowTester | Static analysis for Airflow DAGs, dbt manifests, and airflow.cfg (33 rules) with a unified dependency graph, blast-radius/risk scoring, SARIF/JSON/HTML output, an optional dashboard, and OpenLineage runtime ingestion. |
| StatGuardian | Runtime data contracts in a declarative .sg DSL: schema validation, PSI/KS drift detection, anomaly checks, blocking severity tiers, Pandas/Polars/Delta/Iceberg inputs, and a Rust core. |
| PyBeamGuard | Static pre-flight analysis for Apache Beam, Flink, and Spark pipelines: 10 Beam analyzers plus Flink/Spark checks for hot keys, shuffle, state, checkpointing, windowing, join strategy, cost, and reliability, with YAML rules and CI gating. |
| PyDBTGuard | Pre-deployment dbt test intelligence from manifest.json/run_results.json: risk scoring, historical failure prediction, blast-radius mapping, coverage audits, cost estimates, diagnostics, and opt-in local SQLite/DuckDB/Postgres query checks. |
| PyDependencyCheck | Python dependency intelligence: OSV vulnerability scanning, Git provenance (why/trace), unused-package detection, license compliance, health scoring, CI gating, and signed CycloneDX/SPDX SBOMs. |
| PyDeepMesh | Multi-ecosystem dependency-graph, vulnerability, and attack-surface intelligence (Python, Rust, Node, containers) — manifests, lockfiles, Dockerfiles, Syft image scans, offline/live OSV matching, HTML graphs, SARIF + CycloneDX SBOMs. Static analysis only; never executes your code. |
| PyNetworkIntel | LAN discovery and vulnerability intelligence built on nmap -sV: topology graphs (GraphML/d3), NVD CVE checks, optional SSH config capture, change tracking, NetBox export, Nuclei verification, reports/alerts, and an optional AI summary. |
| PyTaskWireCheck | Local-first, dependency-aware task orchestration: DAG scheduling with bounded concurrency, content-addressed cache, retries, watch mode, SQLite run history, webhooks, and a CLI/TUI/Python API. No server, no daemon. |
| PyStreamXL | Streams multi-sheet .xlsx files row-by-row in constant memory (Rust-backed), with streaming writes/appends, formula/comment extraction, conditional formatting, formula-cycle validation, and an optional REST server. |
| PySynthData | Generates synthetic relational datasets from a schema with referential integrity, constraint enforcement, differential-privacy noise, streaming generation, quality scoring, schema inference, and Parquet/JSON/Pandas export. |
| PyWeatherEnriched | Geocodes a location and fetches real observed historical weather, then engineers climate features (degree-days, rolling stats, anomaly z-scores, offline normals), with backfill, two-tier caching, OSM reverse geocoding, and urban-heat estimates. |
| Repo | What it does |
|---|---|
| PyTokenCalc | Token counting and cost estimation across OpenAI, Anthropic, Google, Cohere, Azure, HuggingFace, Ollama, and custom endpoints — with streaming/incremental counts, pricing overrides and drift warnings, and an optional MCP server. |
| PyInferenceManager | Multi-provider inference orchestrator (Anthropic, OpenAI, Gemini, OpenRouter, Ollama, vLLM) with complexity/privacy/cost routing, hard budget caps, retries/backoff, semantic caching, concurrent and async calls, hardware profiling, and MCP tools. |
| PyStreamMCP | Query planning, source discovery, and context optimization for AI agents — Python SDK, REST API, MCP tool server, multi-endpoint orchestration, framework adapters, and an HMAC-signed webhook event endpoint. |
| PyContextFold | Local-first, structure-aware context optimization for LLMs (Rust core + Python/CLI): folds source code, docs, logs, and structured text into compact representations, with token budgets, repository packing, deterministic profiles, and loss always labeled. |
| PyVectorHound | Diagnostic layer for RAG pipelines: embedding-space and retrieval metrics, BM25/reranker diagnostics, LLM-judge faithfulness checks, root-cause recommendations, drift tracking, batch evaluation, HTML/JSON reports, and opt-in int8 quantization. |
| PyAgentHound | Local-first AI-agent debugger: OpenTelemetry-compatible traces in SQLite, a causal execution graph, deterministic failure rules, ranked root causes, baseline diffs, safe replay, trace-based regression tests, a web UI, and optional Ollama explanations. |
| PyStreamPDF | PDFium-backed PDF parsing with token-budget-aware semantic chunking, dual-tier signed caching, encrypted-PDF permission handling, OCR/validation pipelines, exact or heuristic token counts, and an optional local MCP connector. |
| PyAPICheck | API security discovery and authorization analysis: OpenAPI linting, Schemathesis active testing, spec drift, observed-vs-declared traffic checks, MCP/agent security graphs, behavioral baselining, Cedar policy generation, SARIF output, and remediation plans. |
| PyStreamAI | ML deployment toolkit: canary/A-B routing, cost tracking and budgets, request scheduling, hot model reload, HTTP serving, ONNX Runtime inference, model-registry integrations, and real train/serve abstractions. |
| PyInterviewBot | Voice-first AI interview platform: Rust WebSocket gateway, Python ASR/LLM/TTS engine, and candidate/recruiter client. Runs entirely on local faster-whisper, Ollama, and Piper models; adaptive evidence graph, resume/JD rubrics. Early-stage. |
| OpenAnchor | In-process LLM token intelligence: 6D usage attribution, LangChain middleware, SQLite/in-memory stores, semantic caching, cost governance, and real OpenTelemetry span export. Privacy-safe defaults. |
| PyBlastRadius | Dependency-graph blast-radius analysis, criticality scoring, and failure-cascade simulation, with real parsers for Terraform, Kubernetes manifests, and OTLP traces, plus a CLI, REST API, and Graphviz export. Airflow discovery is still roadmap. |
| Repo | What it does |
|---|---|
| TinyBridge | macOS-native Linux VM runtime on Apple's Virtualization.framework: real Ubuntu guest boot, VirtioFS directory sharing, snapshots, port forwarding, cloud-init seeding, and a daemon/CLI. The macOS backend is real and working; Windows/Linux backends are unimplemented scaffolding. Installed via Homebrew tap, not PyPI. |
| TinyBridge-UI | Native macOS SwiftUI client for TinyBridge: menu-bar control center, VM management windows, live/observed resource metrics, port-forwarding and config validation, and diagnostics with secret redaction. |
| PrismNote | Jupyter-compatible data-science notebook: Rust/Axum backend, React frontend, 15+ code languages, SQL against local engines and 9+ databases (including 8 cloud warehouses), Docker-sandboxed execution, integrated Claude/OpenAI/Ollama assistance, and semantic notebook diff/merge. Installed via PyPI or Homebrew tap. |
| CodeAtlas | Rust-native architecture maps for codebases — deterministic tree-sitter analysis across Python, Rust, Go, JS/TS, Java, PHP, and C#, component grouping, call-graph queries, PR architecture diffs, incremental caching, and Markdown/Mermaid/HTML/MDX/RST output. Optional LLM naming/descriptions. |
| PyPerfGuard | Native benchmark runner and performance-regression guard (Rust): measures wall-clock and resource usage, compares against baselines and Git revisions, reports outliers, tracks local history, integrates with GitHub Actions, and enforces configurable regression policies in any CI — no hosted service. |
A multi-repo, in-development OS project outside the MarTech day job. Most subsystems are currently a deliberate simulation layer — real Rust code with real tests, but in-memory/userspace behavior rather than real hardware I/O or a bootable kernel. Each repo's README states plainly what is real vs. simulated.
| Repo | What it does |
|---|---|
| SHER-KERNEL | Userspace kernel prototype — 40 Rust crates covering a capability-based object model, priority scheduling, tiered memory bookkeeping, driver lifecycle/crash recovery, A/B transactional updates, zero-copy IPC, an AI subsystem, and Linux/POSIX name-level compatibility. Not bootable; simulation-backed. |
| SHER-Graphics | GPU abstraction and rendering — pure-Rust software GPU simulation with a capability-secured runtime, plus a real ash/Vulkan FFI backend verified against MoltenVK; Mesa/OpenGL compatibility is planned. |
| SHER-Display | Native display server, compositor, and window manager: damage-tracking compositor, per-output frame scheduling, scene graph, workspaces, focus/cursor routing, capability-based security, explicit session state, and Wayland/XWayland compatibility seams. Not yet a usable desktop. |
| SHER-INPUT | Canonical, backend-independent input event stream: real Linux/evdev backend, keyboard/pointer/touch/tablet/gamepad event model, global sequencing, capture and synthetic-input contracts, crash detection, and a hardware-free test harness. |
| SHER-Aurora | GNOME-style Rust design system: tokens, typography, Light/Dark/OLED/HDR themes, spring motion, sound events, 24 icons, automated WCAG contrast auditing, and 5 of 17 planned GTK4 widgets rendering real objects today. |
| SHER-Process-Explorer | Evidence-based Linux process investigation: hand-rolled /proc parsing, strace/perf/bpftrace tracing, process intelligence, deterministic "why" findings, log/crash correlation, SQLite history daemon, .deb/.rpm/AUR packaging, and an egui desktop UI. |
| Repo | What it does |
|---|---|
| PyRoboFrames | Rust-backed ML dataloader for robot-learning datasets: native LeRobot v3.0, NumPy/Torch/MLX/JAX outputs, zero-copy frames, Apple Silicon hardware video decode, format conversion (HDF5, RLDS, MCAP, ROS 2 bags), and Hugging Face/S3/GCS loading. |
| PyRoboReplay | Forensic debugging and causal analysis for recorded robot missions: ROS 2 CDR decoding, mission replay/comparison, reality-gap and failure diagnosis, detection gaps, RGB+thermal fusion, Nav2 root-cause checks, and MCAP/Rerun export. |
| PyRoboSimulator | Multi-agent world simulator with real MuJoCo physics, procedural warehouse/parking/urban worlds, RGB/depth/LiDAR/thermal sensors, Gymnasium/PettingZoo adapters, and a FastAPI backend with JWT auth/SSE. Gazebo and Isaac Sim backends are not implemented. |
| PyRoboVision | Robot perception toolkit: Kalman + Hungarian multi-object tracking with occlusion handling, trajectory prediction, depth/3D/LiDAR/BEV occupancy utilities, IMU/GPS fusion, bring-your-own detector, ONNX export, and safety-constraint helpers. |
| PyTerrainMap | Append-only multi-robot terrain store: H3 spatial indexing with confidence decay, HTTP/HTTPS API, terrain/mobility analysis, PostgreSQL persistence, ROS 2 sensor bridge, SLAM/3D reconstruction ops, and Gaussian-splat mapping (partially implemented). |
| SyncFleet | Rust, SQLite-backed peer-to-peer state sync for fleets: mutual-TLS 1.3 with pinned certs and enrollment tokens, version-vector incremental replication, background peer sync with backoff, mDNS discovery, and policy-based conflict resolution (remote_wins/local_wins/last_write_wins/merge_fields). Full CRDTs and mDNS interface binding remain roadmap. |
All repositories · Legacy combined Homebrew tap: homebrew-formulas (superseded by the per-project taps above).
MarTech, data, cloud, AI, and infrastructure stack
MarTech & Customer Data Platforms Adobe RT-CDP, Braze, mParticle, Segment CDP, customer journey orchestration, identity resolution, real-time activation, Hightouch / Reverse ETL
Data Warehousing & Engineering BigQuery, Snowflake, Databricks (PySpark), Microsoft Fabric, Azure Data Factory, Airbyte / Fivetran, dbt, Apache Airflow, Apache Kafka, Apache Flink (PyFlink), InfluxDB, Telegraf, Datadog
Cloud & Data Platforms GCP, Azure, AWS · Google Cloud Storage, Azure Blob Storage / ADLS, Amazon S3, MinIO
AI — RAG & Application Development LangChain, LangGraph, LlamaIndex, Chainlit / OpenWebUI, LangSmith, DSPy, BM25 + vector similarity + reranking, context engineering, Ragas + DeepEval, Pinecone, Weaviate, ChromaDB, pgvector, Qdrant, Milvus
AI — Agent Automation Temporal, Langflow, n8n + Node.js, FastAPI, Firecrawl, Tavily, SerpAPI, Twilio, Claude Agent SDK, OpenAI Agent SDK, Google Agent SDK, Ollama, scheduled and background agents
AI — Autonomous Code Generation Claude Code, Cursor, Replit, Codex, Hermes Agent, OpenCode, Lovable, OpenRouter
MCP & Model Platforms Model Context Protocol, Hugging Face, AWS Bedrock, Azure AI Foundry, GCP Vertex AI (Model Garden)
Infrastructure as Code & Containers Terraform, Docker, Docker Compose, Azure Kubernetes Service (AKS), NGINX Ingress
Identity & Security Entra ID / Azure AD, Azure Key Vault, Azure AI Search, incoming/outgoing guardrails
IoT, Messaging & Robotics Raspberry Pi 5, ESP32, MQTT (Eclipse / RabbitMQ), Node-RED, AWS IoT Core, AWS IoT SiteWise, CoAP, edge processing, industrial IoT
- State the real status. READMEs distinguish shipped, planned, simulated, and known-broken work.
- No fake stubs. A feature is implemented or removed — not left pretending to work.
- Show the evidence. Tests, real browser runs, and confidence labels instead of confident guesses.
- Keep it local-first. Diagnostics should run on your machine, without sending telemetry to a service.
Portfolio-wide audits track confirmed and unverified issues separately. Unverified items stay documented until tested or disproven — they are not dismissed. See the open issues register.
- Email: mullassery@gmail.com
- LinkedIn: georgi-mullassery
- GitHub: github.com/Mullassery
- Location: Bengaluru, India
Always happy to talk customer data platforms, marketing data pipelines, tracking/consent audits, or the tooling that makes them reliable.
If a project here saves you time, a star on the repo helps other engineers find it.