Open to Senior AI / LLM Engineer roles

Taking select AI projects

Sai Ram SanaSenior Agentic AI Engineer

Specialising in RAG & vector search

I design and ship production AI systems (RAG, agent workflows, multi-LLM routing and MCP) with the guardrails and verification that make them trustworthy.

Mississauga, ON · Greater Toronto Area, Canada

Currently Senior Full Stack Developer (AI platform) at Speak AI

See the systems I've built
Illustrative trace of one agent workflow run: route a model, retrieve with RAG, extract, branch, research, act, verify.
Years shipping software
9+
LLMs routed through one model registry
45+
Transcription languages supported
137
Live translation languages
51

// about

Engineer at the intersection of AI, code and revenue

I'm a senior engineer with 9+ years of experience shipping production software. Since 2023 I've been building the AI platform at Speak AI: RAG and semantic search, a branching agent workflow engine, multi-LLM routing, multimodal audio and video AI, MCP tools, and real-time transcription and translation.

I build with agents too, using multi-agent research with adversarial verification, read-only production diagnostics agents and parallel agent worktrees, so delivery gets faster without lowering the bar. I'm looking for Senior AI Engineer, LLM Engineer and Agentic AI roles, and I take on select AI automation projects.

What I do

  • RAG & vector search over any data
  • Branching agent workflows & MCP tools
  • Multi-LLM routing & multimodal AI
  • Real-time speech & live translation
  • AI sales automation & prospecting
  • Lead teams, review code, mentor devs

Education

  • Post-Graduate Diploma, Business and Information Systems Architecture

    Fanshawe College · London, Ontario, Canada

    President's Honours
  • Bachelor of Technology, Computer Science Engineering

    Jawaharlal Nehru Technological University · India

// systems i've built

Production AI, end to end

The AI systems I've designed and shipped. Each card shows the problem, the architecture and the key decision behind it.

Retrieval · RAG

RAG & Semantic Search over Conversations, Docs and Websites

Teams have hundreds of hours of recordings, documents and web content, and need answers they can trust, not guesses.

  1. Ingest
  2. Chunk
  3. Embed
  4. Vector store
  5. Filter + retrieve
  6. Grounded answer

Key decisionFilter by metadata before vector search, and never answer without citations. It trades a little recall for answers users can verify.

How it's built
  • Transcripts, documents and crawled websites are chunked with a recursive text splitter and embedded with OpenAI text-embedding-3-large (1024-d).
  • Conversation vectors live in Pinecone. Knowledge-base vectors use MongoDB Atlas Vector Search, with a local similarity fallback.
  • Metadata filters (folder, speaker, tags, date, custom fields) narrow retrieval before top-k search, so answers stay scoped to the right data.
  • Answers cite their sources, and users can open the exact media, document or page each claim came from.
  • Knowledge bases refresh incrementally: sitemap lastmod diffing re-embeds only changed pages, with upload-before-delete swaps and a single re-embedding pass.
  • Vector lifecycle tooling archives and cleans orphaned embeddings when media is deleted or archived.

Chat with your data across media, documents and websites, with cited sources and indexes that stay fresh without full re-indexing.

  • OpenAI Embeddings
  • Pinecone
  • MongoDB Atlas Vector Search
  • LangChain
  • Node.js
  • TypeScript

Agents · Orchestration

Branching Agentic Workflow Engine + Visual Builder

Customers wanted AI to act on every new recording (extract, decide, research, notify, sync) without writing code.

  1. Trigger
  2. AI step
  3. Condition
  4. Branch A / B
  5. Actions
  6. Run history

Key decisionRuns are resumable graphs, not one long LLM call. Each step checkpoints, so a failed step re-runs on its own instead of repeating the whole (paid) run.

How it's built
  • A graph (DAG) runner executes steps in dependency order, with nested if/else condition branches that re-join at a shared step.
  • Step types include AI prompt, structured field extraction, web research, translation, third-party actions, outbound webhooks and notifications.
  • Triggers can be events (recording received, field updated with AND/OR matching, AI chat completed), schedules, inbound webhooks or integration events, and several can be OR-combined.
  • Runs pause on long work (media processing, audio extraction, translation), resume idempotently from the frontier, and can re-run just the failed step.
  • Guardrails include a cascade-depth limit against automation loops, per-run config snapshots, save-time graph validation, and test runs with no billing or side effects.
  • The visual canvas (React Flow) supports lane-based branches, variable tokens between steps, a step wizard, test drawer and step-by-step run timeline.

No-code AI agents that branch on what the model finds, so every recording can be routed, enriched and acted on automatically.

  • TypeScript
  • Node.js
  • MongoDB
  • React 19
  • React Flow
  • Zustand
  • TanStack Query

LLM Platform

Multi-LLM Model Registry & Router (with BYOK)

Model logic was duplicated across the codebase, and one hard-coded default could push every workspace into the same provider rate limit.

  1. Request
  2. Registry
  3. Resolver
  4. Provider driver
  5. Stream
  6. Usage + billing

Key decisionOne registry is the single source of truth for every model. Adding or retiring a model is a data change, not a code change, and every routing choice is logged with its reason.

How it's built
  • A single shared registry defines 45+ models across OpenAI, Anthropic Claude, Google Gemini, xAI Grok and GLM, with lifecycle state, pricing and capability flags (vision, thinking, native audio/video, temperature support).
  • One resolver chooses the model for every request and records an auditable reason, covering explicit requests, workspace preferences, defaults, safe substitution of retired models, and rerouting media to capable models.
  • Bring-your-own-key (BYOK) routes turns onto a customer's own OpenAI, Anthropic, OpenRouter or Vertex keys, and bills at the rate of whichever key actually paid.
  • Provider drivers handle streaming, tool calls, retries with backoff on 429s, and quirks such as models that reject custom temperature.
  • Structured JSON-schema extraction works consistently across all providers.

New models launch by adding one registry entry, and every routing decision is explainable, cost-aware and resilient to provider limits.

  • Claude
  • GPT
  • Gemini
  • OpenRouter
  • TypeScript
  • Shared npm package

Multimodal AI

Audio, Video & Image Understanding on Gemini / Vertex AI

Transcripts miss what's on screen and how things are said, so users wanted the model to actually watch and listen.

  1. Media
  2. Eligibility
  3. Stage (S3 → GCS)
  4. Vertex region
  5. Gemini
  6. Guardrails

Key decisionEvery request is checked for eligibility and priced before it's sent. Cost stays predictable, and unsupported media fails fast with a clear reason.

How it's built
  • A pure eligibility function checks container and codec against provider allowlists and plan duration caps, and returns machine-readable refusal reasons.
  • Small files go inline; large files stream from S3 into the Gemini Files API or Google Cloud Storage for Vertex. YouTube URLs and photo carousels are supported natively.
  • Every request is priced before it runs using per-second audio and per-frame video token models, and the same quote drives the credit check and the bill.
  • Vertex regions follow the customer's data-residency region, with support for customer-owned GCP projects and no silent fallback where compliance forbids it.
  • A bias guardrail stops the model from judging protected characteristics such as accent, race or age from audio, video or images.

Users can ask questions about what happens in a video, not just what was said, with predictable cost and data residency.

  • Gemini
  • Vertex AI
  • Google Cloud Storage
  • AWS S3
  • FFmpeg
  • TypeScript

MCP · Tool Use

MCP Server: Let Claude & ChatGPT Operate the Platform

Power users wanted to drive their media, insights and automations from inside Claude and ChatGPT.

  1. Claude / ChatGPT
  2. MCP tools
  3. Validate
  4. API
  5. Result

Key decisionAssistants get validate and test tools next to build tools, so they verify a workflow before switching it on instead of acting blind.

How it's built
  • Built MCP tools spanning media, transcripts, AI analysis, automations, webhooks and run monitoring, published as an npm MCP server.
  • A build_automation tool lets an assistant design complete branching workflows. A graph compiler turns the editor's branch tree into the server's graph format while enforcing depth, step and type limits.
  • Tools to validate, describe, test-run and inspect automations give the assistant a verify-before-act loop.
  • Guided inbound-webhook tools return the receive URL, sample payload and ready-to-use variable tokens.
  • Backed by extensive unit, smoke and live-harness tests.

Anyone can say "when a sales call lands, extract the objections and post them to Slack" in Claude and get a working automation.

  • Model Context Protocol
  • TypeScript
  • Zod
  • Vitest
  • Claude
  • ChatGPT
More systems (4): Agents, Real-time Speech, Sales AI, Cross-platform Apps

Agents · Grounding

Grounded Web Research Step for Agents

Workflow agents needed fresh, outside context (company news, facts, references) without hallucinating.

  1. Transcript
  2. LLM query
  3. Search API
  4. Score filter
  5. Cited synthesis

Key decisionWeb results are untrusted input. They're relevance-filtered, cited, and never allowed to reach the prompt-template layer.

How it's built
  • An LLM writes the search query from the transcript, and can return NO_QUERY to skip instead of fail.
  • Pluggable providers (Tavily, Perplexity) are chosen per step. Results are filtered by relevance score and capped, keeping only the best sources.
  • The model writes its own answer with citations. Third-party answers are passed in as unverified context only.
  • Prompt-injection defence escapes template syntax found in web content, and the step fails before spending when a key is missing.
  • Sources and citations are stored with the result for audit.

Agents enrich every conversation with verifiable outside context, safely.

  • Tavily
  • Perplexity
  • LLM
  • TypeScript

Real-time Speech

Live Transcription & Live Translation Platform

Meetings and recordings needed captions and translations while they happen, not after.

  1. Mic / meeting bot
  2. WebSocket
  3. Streaming ASR
  4. Sentences
  5. Translate
  6. Fan-out

Key decisionTranslate at sentence boundaries, not word by word. A short delay buys translations that actually read correctly.

How it's built
  • A streaming gateway (Socket.IO plus raw WebSocket) takes audio from browsers, native apps and meeting bots, with diarization and interim and final results.
  • Streaming ASR on Deepgram chooses the best model per language (48 live languages), and audio is streamed to S3 via parallel multipart upload as it's transcribed.
  • Real-time FFmpeg extraction pulls audio from incoming video chunks, with memory monitoring.
  • A sentence-boundary detector (punctuation, CJK-aware, or a pause) drives live translation into 51 languages, fanned out per room with late-joiner backfill.
  • Batch speech-to-text spans 4 engines (AWS, Azure, AssemblyAI, Deepgram) with automatic fallback and 137 languages. Meeting bots auto-join from Google and Outlook calendars.

Live captions and translations for multilingual meetings, and a single speech layer across four ASR providers.

  • Node.js
  • Socket.IO
  • WebSockets
  • Deepgram
  • AssemblyAI
  • FFmpeg
  • AWS S3

Sales AI · Voice

AI Prospecting Agent → Personalised Voice-Agent Demos

Sales wanted every prospect to see a demo built for them, without manual prep.

  1. Apollo list
  2. Enrich
  3. Scrape site
  4. Embed KB
  5. Voice agent
  6. Outreach

Key decisionRefresh each prospect's knowledge base incrementally instead of rebuilding it, so agents stay current at low cost.

How it's built
  • A scheduled poller pulls new contacts and organisations from Apollo lists within a rate-limit budget, with locking, dry-run and admin controls.
  • Each prospect's website is crawled and embedded into a knowledge base, and a personalised demo voice agent is generated automatically.
  • Knowledge bases stay current with incremental re-scraping and re-embedding.
  • Voice previews are generated with OpenAI TTS, and ElevenLabs voice models are configurable per agent.

Every prospect gets a voice agent that already knows their business, built automatically from the CRM list.

  • Apollo
  • RAG
  • LiveKit
  • ElevenLabs
  • OpenAI
  • Node.js
  • Next.js

Cross-platform Apps

Native iOS/Android & Desktop Recorder Apps

Users needed reliable recording anywhere (phone, browser, desktop), even offline.

  1. Record
  2. Native module
  3. Offline queue
  4. S3 upload
  5. AI pipeline

Key decisionRecord locally first and upload in the background with resumable retries, so a dropped connection never loses a recording.

How it's built
  • Expo / React Native app with custom native modules in Swift and Kotlin for AAC recording with metering and an Android documents provider, published to the Play Store.
  • Studio UI with live waveform, takes, haptics, an offline library, resumable auto-upload and crash-safe recording.
  • Deep links (Universal Links / App Links), 6-digit pairing, Apple and Google sign-in, and biometric lock.
  • Embeddable web recorder with live transcription and a white-label theming engine that validates custom CSS.
  • macOS desktop recorder with Sparkle auto-updates and meeting mode, plus an earlier Angular + Capacitor mobile app.

One recording experience across web, iOS, Android and desktop, feeding straight into the AI pipeline.

  • React Native
  • Expo
  • Swift
  • Kotlin
  • Capacitor
  • Angular
  • React

// agentic ai engineering

Agents that plan, act, verify and ship

I treat LLMs as components in engineered systems. Every agent I ship has a clear goal, scoped tools, memory, guardrails and a verification step. I build software the same way: with agents, under the same discipline.

  1. 01

    Plan

    Agents decompose the goal and pull context from code, docs, data and CRM.

  2. 02

    Act

    Scoped tools and MCP servers write code, query data, enrich leads and draft content.

  3. 03

    Verify

    Tests, schema checks, an adversarial second model and read-only production evidence.

  4. 04

    Decide

    Branch, retry, escalate to a human or ship, based on confidence and rules.

  5. 05

    Learn

    Runs are logged and failures become evals and guardrails for the next run.

↻ outcomes feed back into the next run

How I build software with agents

agent.01

Multi-agent research squad with an adversarial verifier

14 specialist agents (architecture, realtime, audio pipeline, native platform, migration, compliance and more) work under a shared evidence contract: every claim needs a dated source. A verifier agent tries to disprove each finding, and a synthesizer builds the report only from what survives. I used this to choose a cross-platform architecture from 8 candidates.

agent.02

Read-only production diagnostics agents

On-call agents inspect production data, logs and Sentry through MCP, match symptoms against a catalogue of known failures, and report verdict, cause and next action. They are designed never to change anything.

agent.03

Parallel agents in isolated git worktrees

Features, security hardening and PR reviews run as separate Claude Code agents in their own worktrees, so work proceeds in parallel without collisions.

agent.04

Verification-first delivery

Design docs are validated with read-only production queries, reviewed adversarially, then shipped behind feature flags with parity tests, golden outputs and kill switches.

agent.05

Guardrails around coding agents

Pre- and post-tool hooks block secrets and production writes, run test, comment and duplication checks, and gate PR creation on review. Humans always merge.

Models I build with

  • Claude (Opus / Sonnet / Haiku)
  • GPT-4o / GPT-5 family
  • Gemini (Vertex AI)
  • Grok & GLM via OpenRouter
  • Whisper / Deepgram / AssemblyAI
  • OpenAI TTS & ElevenLabs

Practices

  • Retrieval-augmented generation (RAG)
  • Vector embeddings & semantic search
  • Model Context Protocol (MCP)
  • Multi-agent orchestration
  • Tool / function calling
  • Multi-LLM routing & BYOK
  • Multimodal (audio / video / image)
  • Structured output extraction
  • Prompt-injection defence
  • Bias & safety guardrails
  • Cost-aware token pricing
  • Human-in-the-loop design

// ai sales automation

End-to-end AI sales pipelines

I build revenue engines, not just lists. My sales automations connect data providers, AI research agents and CRM, so the right account gets the right message at the right time, automatically.

  1. 1

    Source

    Pull target accounts and signals from Apollo, Clay, LinkedIn and intent data.

  2. 2

    Enrich

    Waterfall enrichment for firmographics, tech stack, contacts and verified emails.

  3. 3

    Research

    AI agents read websites, news and job posts and build a per-account brief.

  4. 4

    Score

    Agentic scoring against each ICP profile decides priority and route.

  5. 5

    Engage

    Personalised, AI-written sequences, or a custom voice-agent demo per prospect.

  6. 6

    Route

    Replies classified by AI, synced to CRM and handed to the right rep.

Sales tools I work with

  • Clay
  • Apollo.io
  • AI Ark
  • HubSpot
  • Instantly
  • Smartlead
  • LinkedIn Sales Navigator
  • n8n
  • Zapier
  • Make
  • Claude
  • GPT

Agentic scenarios I've automated

  • New Apollo prospect→site crawled & embedded→personalised voice-agent demo
  • Inbound lead→instant AI research→ICP fit score→routed to rep or nurture
  • Hiring / funding signal→account brief→personalised outbound sequence

// experience

Career timeline

  1. Senior Full Stack Developer (AI Platform)

    Speak AI Inc.· Mississauga, Canada

    Jul 2023 — Present

    Building the AI platform of a conversation-intelligence product: RAG, agent workflows, multi-LLM routing, multimodal AI, MCP, real-time speech and cross-platform apps. I also lead development and code review.

    + show highlights (8)
    • Designed and built the AI systems shown above: RAG and semantic search, the branching agent workflow engine, the multi-LLM registry and router, multimodal Gemini analysis, MCP tools and the grounded web-research step.
    • Built a real-time transcription gateway (WebSocket/Socket.IO, streaming ASR, 48 live languages) and live translation into 51 languages.
    • Built an Apollo-driven prospecting agent that auto-generates a personalised voice-agent demo from each prospect's website.
    • Shipped a native iOS/Android recorder (Expo, Swift, Kotlin), an embeddable recorder with white-label theming, and a macOS desktop recorder.
    • Integrated meeting bots with Google/Outlook calendar auto-join, plus Google Drive, OneDrive, SharePoint, Slack, ClickUp and Zoom workflow actions.
    • Implemented Python sentiment analysis (+25% insights) and deployed NER models on autoscaling AWS ECS behind an ALB.
    • Integrated Stripe, PayPal and Paddle billing, OAuth 2.0/SSO, multi-region S3 data residency, and SSRF-safe outbound requests.
    • Cut latency by 40% with Redis caching. I lead development cycles, code reviews and mentoring.
    • TypeScript
    • Node.js
    • Python
    • React
    • React Native
    • Angular
    • AWS
    • GCP / Vertex AI
    • MongoDB
    • Pinecone
    • Claude
    • GPT
    • Gemini
    • MCP
  2. Client Service Controller (Individual Customer)

    Canada Life· London, Canada

    Mar 2023 — Jul 2023

    Automated insurance policy workflows with APIs and event-driven microservices.

    + show highlights (4)
    • Analysed business processes to find opportunities for automation.
    • Designed an API that automates processing of individual client and Ingenium policies.
    • Implemented Kafka producers and consumers in microservices for real-time data flow.
    • Led MongoDB version upgrades and migrations across platforms.
    • Kafka
    • Microservices
    • MongoDB
    • Backbone.js
    • JUnit
  3. Full Stack Developer

    Justkapture Innovations· London, Canada

    Apr 2021 — Feb 2023

    Built marketing-tech products that turn social and on-site engagement into sales.

    + show highlights (6)
    • Built FOMO, an app that tracks end-user interactions on web applications to drive sales.
    • Built Justkapture, which gathers tagged content from social media sources to improve customer sales.
    • Delivered full stack features with Node.js, Express, MongoDB, Angular, TypeScript and AWS.
    • Designed PostgreSQL schemas with complex stored procedures and functions.
    • Implemented OAuth 2.0, OAuth 1.0a and JWT for third-party vendor integrations.
    • Developed APIs to monitor and manage AWS Lambda services.
    • Node.js
    • Angular
    • React
    • TypeScript
    • PostgreSQL
    • AWS Lambda
    • Java
    • Spring Boot
  4. Associate Developer

    Cognizant Technology Solutions· Chennai, India

    Dec 2018 — Apr 2021

    Built enterprise investment-data services for trading desks.

    + show highlights (5)
    • Designed and developed the Investment Data Service Controller (IDS), which aggregates Bloomberg, Flex and other feeds for traders.
    • Automated manual checks, saving about 15% of support team effort annually (~$120,000).
    • Built an internal Spring Boot testing library for TWS (IBM Scheduler) QA.
    • Used Adobe Experience Manager (AEM) for dynamic content delivery.
    • Applied OOAD, UML, Spring MVC, AOP and Hibernate ORM.
    • Java
    • Spring
    • Hibernate
    • MSSQL
    • AEM
    • JUnit
    • Mockito
  5. Software Engineer

    Medoske Technology Solutions· Hyderabad, India

    Jun 2017 — Dec 2018

    Built healthcare web applications for clinics and practitioners.

    + show highlights (3)
    • Built doctor appointment and patient case-sheet apps with React and MobX.
    • Built a dental analysis platform that improves diagnostic efficiency.
    • Built the Redheal Care platform with React and Next.js, with real-time communication and live updates.
    • React
    • Next.js
    • MobX
    • JavaScript

// skills

The toolkit

LLMs & Agents

  • Claude
  • GPT / OpenAI
  • Gemini / Vertex AI
  • OpenRouter
  • Multi-agent orchestration
+4 more
  • Tool calling
  • MCP servers
  • Claude Code
  • Structured outputs

RAG & Retrieval

  • Vector embeddings
  • Pinecone
  • MongoDB Atlas Vector Search
  • LangChain
  • Chunking strategies
+2 more
  • Metadata filtering
  • Web search grounding (Tavily, Perplexity)

Speech, Voice & Multimodal

  • Deepgram
  • AssemblyAI
  • AWS Transcribe
  • Azure Speech
  • Live translation
+4 more
  • LiveKit voice agents
  • ElevenLabs / OpenAI TTS
  • Gemini audio & video
  • NER & sentiment

Sales Tools & Automation

  • Clay
  • Apollo.io
  • AI Ark
  • HubSpot
  • Outbound sequencing
+3 more
  • Lead enrichment
  • ICP scoring
  • n8n / Zapier / Make

Backend

  • Node.js
  • TypeScript
  • Express
  • Python
  • Flask
+5 more
  • REST & WebSockets
  • Socket.IO
  • Queues & schedulers
  • Redis
  • Kafka

Frontend & Mobile

  • React 19
  • Next.js
  • React Native / Expo
  • Swift & Kotlin modules
  • Angular
+4 more
  • Vue.js
  • React Flow
  • Zustand / TanStack
  • Capacitor

Cloud & Data

  • AWS (S3, ECS, Lambda, CloudFront, Cognito, SES, SSM)
  • GCP (Vertex AI, GCS)
  • Cloudflare
  • Docker
  • MongoDB
+2 more
  • PostgreSQL
  • FFmpeg

Foundations

  • Java
  • Spring Boot
  • Hibernate
  • OAuth 2.0 / SSO
  • Stripe / Paddle / PayPal
+3 more
  • Design patterns
  • Testing (Vitest, Playwright, JUnit)
  • Agile

// writing

Notes on agentic AI

// work with me

Need AI that works in production?

Alongside full-time roles, I take on a small number of advisory and contract engagements, helping teams ship AI that's reliable, safe and worth the cost.

AI agents & workflow automation

Agents that plan, act and verify, wired into your tools, with human approval where it matters.

RAG & knowledge assistants

Search and Q&A over your docs and data, with citations, access control and fresh indexes.

MCP & multi-LLM integration

Connect Claude or ChatGPT to your product through MCP, with model routing across providers.

AI architecture review

A focused review of an existing AI feature: reliability, cost, safety, evals and a prioritised fix list.

How it works

  1. 01

    Discovery call

    30 minutes, free. Your goal, data and constraints.

  2. 02

    Scoped proposal

    Clear scope, milestones and timeline before any work starts.

  3. 03

    Build & verify

    Weekly demos, tests and guardrails, so you see progress early.

  4. 04

    Handover

    Docs, runbook and evals, so your team owns it.

Book a discovery call

// let's build

Need agents that actually ship results?

Whether it's an agentic engineering workflow, an AI sales automation or a full stack AI product, let's talk.