Ismail Kattakath

Principal engineer for Takeoff — a multi-tenant AEC estimation AI / digital-takeoff SaaS on AWS Canada Central.

• Own the backend & cloud platform: Python 3.12 monorepo, FastAPI on AWS Lambda, multi-stack AWS CDK
• Multi-tenant security: dual Cognito pools, token authorizers, WAF, Aurora Serverless v2 + RDS Proxy (IAM auth)
• SaaS billing on Stripe: free/trial/paid lifecycle, webhooks, reconciliation, AI usage metering & overage billing
• Applied-AI platform for the document-AI worker fleet: an OpenAI-compatible LiteLLM inference gateway (built, then a disciplined teardown) and an AI cost-governor with a fail-closed kill-switch; diagnosed a by-construction-invisible RAG failure — provider/dimension drift silently collapsing vector search to keyword-only scoring while self-reporting success — closed with a CI parity test making it structurally impossible to reintroduce
• Cross-cutting platform reliability across a 20+ Lambda monorepo: a serverless Alembic migration-runner (PG advisory-lock serialized, offline SQL dry-runs), a fleet-wide async-worker idempotency contract for exactly-once effect, and structured observability (Lambda Powertools + X-Ray correlation IDs)
• CI/CD via GitHub Actions with GitHub OIDC (no stored keys), two-speed hotswap/full-CFN deploys, staging-as-required-gate, Alembic migration runner, and Bruno smoke tests; agent-fleet guardrails treating the LLM dev-agents as untrusted actors

Principal engineer across two workstreams: GenAI infrastructure (provider-abstracted inference, LLM pipelines, MCP and RAG implementations), and DontSell.ai — an event-driven, Pub/Sub-based autonomous-outbound (AI-SDR) platform built solo as a fleet of GCP Cloud Run microservices.

• Architected provider-abstracted GenAI inference — Strands Agents behind one LLM_PROVIDER across Vertex AI (prod), local Ollama (offline dev), and a LiteLLM proxy — so identical generation code runs free-local or managed-cloud via one switch
• Built DontSell.ai solo: an event-driven, Pub/Sub-coordinated AI-SDR platform as a fleet of serverless GCP Cloud Run microservices — signal ingestion, explainable versioned account scoring, ICP synthesis, and outreach-draft generation
• Hardened the message pipeline for at-least-once delivery — HMAC-verified webhook ingestion, idempotent upserts, and poison-message ACK semantics across every worker
• Designed the human-in-the-loop (HITL) outreach lifecycle with compliance-by-design: a 6-state approval-gated draft queue, immutable AI-disclosure, and anti-spam governance baked into the schema
• Productionized agentic-AI pipelines on robust cloud infrastructure with scalability, observability, security, IaC, and CI/CD

Worked on Canada's leading digital mental-health platform (homeweb.ca), delivering EFAP and CBT solutions to 100k+ employees nationwide.

• Inherited and revitalized an abandoned, undocumented MEAN-stack application into a maintainable, production-ready platform with 99.5% uptime
• Rebuilt the authentication system with OAuth 2.0 PKCE for SSO integration with Microsoft Identity Platform, supporting 25,000+ users
• Collaborated with federated vendors to configure and troubleshoot SAML 2.0, enabling seamless login across 5+ partner organizations
• Led the complete frontend migration from AngularJS to Next.js on framework EOL, preserving UI/UX and API compatibility
• Established DevOps practices — automated testing, CI/CD (cut deployment from 4 hours to 20 minutes), containerization, monitoring, and backups
• Led infrastructure migration from DigitalOcean to AWS (EC2, RDS, S3, CloudFront, VPC) using Terraform, achieving zero-downtime deployment

Concurrent short-term contract (alongside Homewood) for scheduled maintenance of business-critical legacy IBM systems, executing upgrades and infrastructure improvements with zero downtime.

• Executed zero-downtime upgrades of 3 IBM OpenPages GRC and 3 Cognos BI environments (7.3→8.0.0.2, Cognos 10→Analytics 11) with full business continuity
• Configured an IBM AIX load-balanced HA cluster and automated the upgrade process end-to-end with Bash

Built a complete hospitality software suite (Brilliant Rewards) — dashboard, mobile and web apps — with loyalty programs, marketing automation, booking integration, and analytics.

• Led development of a scalable, containerized microservices architecture with discoverable OpenAPI-spec'd services
• Introduced development standards (coding conventions, linting, code reviews, incident management) reducing production incidents by 65%
• Delivered multiple client applications (AngularJS web + cross-platform mobile) while managing sprint planning and stakeholder requirements end-to-end

Built a school MIS (RM Integris) covering student management, grade tracking, attendance, parent portals, and performance analytics.

• Implemented an OpenAPI/Swagger bridging layer to modernize legacy SOAP APIs for RESTful clients
• Prototyped an AngularJS interface to decouple the presentation layer from a legacy Struts/JSP architecture
• Maintained load-balanced Linux VM clusters hosting distributed web, application, and database tiers

Worked on an adaptable cloud ERP (Posibolt) serving retail businesses across global industry segments.

• Created UI mockups from SRS specifications and converted approved designs into responsive HTML/CSS templates
• Developed integration test cases and implemented browser automation using Selenium WebDriver
• Authored user documentation including help articles, user guides, FAQs, and product website content

Local-First RAG Platform & Evidence-Gated Resume Pipeline

A zero-key, loopback-only RAG stack (published as the MIT flake nix-local-rag): launchd Postgres + pgvector with an in-database embed() → vector(768) function (nomic-embed-text via Ollama) and HNSW cosine retrieval. Powers a ~300-document career-evidence store that gates an AI job-application pipeline — every resume claim must be corroborated by mined, git-attributed evidence before it ships.

• In-DB embeddings — retrieval without any external API key; nothing leaves the machine
• Evidence-gated tailoring with adversarial truthfulness cross-checks against the store
• Dogfooded: extracted as a standalone flake, consumed back as an input

MCP Gateway Fleet (Local + Public OAuth Endpoint)

A declarative ~20-server Model Context Protocol gateway: one mcp-proxy launchd agent on loopback hosting Nix-pinned and stdio MCP servers for every local AI client, with secrets fetched from the macOS Keychain at launch (never in argv or the store) and known upstream failure modes pinned. A second OAuth-gated proxy publishes a chosen server subset through a Cloudflare Access managed-OAuth tunnel; a self-hostable video-MCP stack (HyperFrames) ships the same fail-closed pattern as OSS.

• RCE-surface servers deliberately excluded from the shared gateway (per-client stdio only)
• Public exposure is allowlist-only behind managed OAuth — fail-closed by default
• Deep MCP internals fluency: transports (stdio/SSE/streamable HTTP), MCP-OAuth, server packaging

Supervised Agentic-Engineering Harness & Automation Fleet

Production discipline for LLM dev-agents: a crash-safe hook supervisor (loop-breaking, incident logging) wrapping evaluation gates so autonomous coding agents cannot land broken configs; a 19-specialist multi-agent development architecture with team dispatch; and a declarative browser-automation fleet — a dedicated CDP-driven Chromium with headed-auth/headless-work modes and Playwright sessions kept encrypted in the macOS Keychain, treating agents as untrusted actors throughout.

• Agent guardrails as infrastructure: gitflow gates, prompt-based tool policies, fail-closed evaluation hooks
• Reliable computer-use path: CDP/Playwright attach over extension-based control, with session credentials never plaintext on disk
• Declarative nix-darwin/NixOS fleet management (macOS + Raspberry Pi + VMs) as the substrate

Diffusion MLOps on Rented GPUs

Production-grade provisioning for open-weight diffusion pipelines: ComfyUI node graphs (Flux-2, Qwen Image Edit, LTX audio+video I2V; 17–104 nodes with tuned samplers, stacked LoRAs, GGUF-quantized encoders) wrapped in reproducible one-click provisioning across two rented-GPU providers (Vast.ai, RunPod) — sha256 fail-closed model-integrity gates, pinned custom-node commits, and preflight credential probes so a broken instance never reports ready.

• Fail-closed integrity gating: hash mismatch or missing model fails provisioning, never silently degrades
• Cross-provider parity debugging (torch-family conflicts, provisioner lifecycle deadlocks, CDN redirect quirks)
• Published as a standalone provisioning flake (nix-vast-provision), dogfooded from the fleet