Senior Frontend & AI Engineer — Toronto

Deepak Singh Kandari

I build agentic AI products — and the frontend craft around them. One lineage of engineering, dispersed across 27 projects.

View the work Get in touch
The lineage

From Silicon to Sentience

Every project below stands on this history — the transistor to the autonomous agent.

The Transistor
1947The TransistorSilicon switches a current. Computing begins.
The Microchip
1971The MicrochipThousands of switches on a fingernail of silicon.
The GPU
1999The GPUParallel math — the engine that would train minds.
Neural Nets
2012Neural NetsAlexNet. Machines start to see.
Transformers
2017TransformersAttention is all you need. Language unlocks.
Agents
2023AgentsModels that act — reason, call tools, ship work.

Seventy years of engineering converge into four clusters of work.

C1Foundations

I learned the model from the inside out.

Before touching an API, I built the primitives by hand — pre-training, alignment, reinforcement learning, and inference at the memory level. This is the credibility floor everything else stands on.

3 projects · disassembled from the core

ML Portfolio object

ML Portfolio

Three systems from scratch in raw PyTorch — a GPT with RLHF/DPO, a PPO game agent, and a feedback-learning RAG. No wrappers.

PyTorchRLHF / DPOPPO + GAEBM25 + FAISS
View details Code
Parameter Golf object

Parameter Golf

OpenAI's efficiency challenge — the best language model that fits in 16 MB and trains in 10 minutes on 8×H100. QAT, the Muon optimizer, scaling laws.

QuantizationMuonScaling Laws16 MB
View details Code
ACMA object

ACMA

Adaptive Context Memory Allocator — joint prompt-compression + mixed-precision KV-cache + token eviction as a knapsack under one GPU budget. Research-grade.

KV-CacheFP16→INT2LLMLingua-2Knapsack
View details Code
C2Infrastructure

The plumbing that makes AI production-grade.

Models are the easy part; the system around them is the job. Cost-routed serving, RAG with continuous evaluation, multi-agent orchestration, on-prem fine-tuning, and AI woven into the developer workflow — all on consumer hardware.

11 projects · disassembled from the core

LLM Gateway object

LLM Gateway

One OpenAI-compatible API across five providers — cost-routing, semantic caching, and circuit-breaker failover with Prometheus/Grafana observability.

HonoRedisSemantic Cache5 Providers
View details Code
RAG Eval Engine object

RAG Eval Engine

A RAG pipeline with evaluation baked in — hybrid BM25 + dense via RRF, adaptive retrieval, and a 5-metric LLM-as-judge so quality regressions show up instantly.

QdrantRRFLLM-as-Judge68 Tests
View details Code
AgentHire object

AgentHire

Five agents, one pipeline, zero API cost — each agent its own MCP server, coordinated with LangGraph and running 100% on local inference.

LangGraphMCPOllamaTurborepo
View details Code
Enterprise Playground object

Enterprise Playground

Trained a 14B local model on real enterprise UIs so PMs generate prototypes at $0/token — the full data → QLoRA fine-tune → serve loop on one GPU.

QLoRADual-ModelChromaDB$0/token
View details Code
Git Sentry object

Git Sentry

An AI git guardian — four local agents block bad commits at $0, and CI escalates only complex files to Claude, cutting review spend ~75%.

OllamaDual-Phase SecurityGitHub Actions~75% Cut
View details Code
Context Engineering Academy object

Context Engineering Academy

Teaching the full AI-engineering stack through interactive, quality-scored playgrounds — context, evals, agents, MCP, and observability.

6 Academies70+ ModulesLive Scoring
View details CodeLive
Harvest object

Harvest

Batch AI tasks across free model quotas and local models — fan work out, recover from errors, pay nothing.

BatchMulti-QuotaLocal LLM$0
Open project Code
Composio Hiring Pipeline object

Composio Hiring Pipeline

An AI hiring pipeline + MCP bridge — multi-tool orchestration with error recovery, built as platform engineering.

ComposioMCP BridgeOrchestration
Open project Code
Claude Dashboard object

Claude Dashboard

Real-time monitoring for the Claude Code ecosystem — see everything, miss nothing.

WebSocketCanvas 3DReal-time
View details Code
Claude Pilot object

Claude Pilot

Mission control for Claude Code — analytics, agent management, and MCP control in one command.

MCPTypeScriptnpx
View details Code
Tech Deep Dive object

Tech Deep Dive

Deep technical breakdowns of the projects and the technologies behind them — because understanding compounds.

MDX54 TechnologiesInteractive
View details Code
C3Agentic Products

Agents you can actually ship in a bank.

Grounded RAG with verifiable citations, human-in-the-loop approval, and a tamper-evident audit trail. Governed copilots for regulated domains — the same architecture built as a Python backend and as a zoneless Angular frontend.

6 projects · disassembled from the core

Meridian object
Flagship

Meridian

An Angular 21, fully zoneless, signals-first agentic banking-policy copilot — verifiable citations, human-in-the-loop, FAPI 2.0 / DPoP. The BMO Lumi pattern, in Angular.

Angular 21SignalsZonelessDPoP
View details CodeLive
ALPHA Advisor object
Flagship

ALPHA Advisor

A LangGraph agentic-RAG wealth-advisory copilot — interrupt-driven human-in-the-loop, knowledge-graph retrieval, governance, and a hash-chained audit trail.

LangGraphGraphRAGHITLAzure OpenAI
View details CodeLive
Beacon object

Beacon

A 65 KB, zero-dependency agentic workspace — grounded citations, human-approved actions, span feedback. Built for air-gapped enterprise.

React 19Cohere v2SSE65 KB
View details Code
Nosce object

Nosce

Know thyself — an AI that builds a knowledge graph of how you think, then reflects it back.

Knowledge GraphNLPAnalytics
View details Code
Sentinel L4 Trust Engine object

Sentinel L4 Trust Engine

A trust-scoring engine for Stripe agentic commerce — the missing L3 → L4 layer that lets agents transact safely.

Agentic CommerceTrust ScoringStripe
Open project Code
BrainRot Index object

BrainRot Index

An AI that scores how cooked your thinking is — six cognitive dimensions, scored by Gemini.

Gemini6 DimensionsScoring
Open project Code
C4In Market & Craft

Built, billed, shipped — and made beautiful.

Not prototypes. Live products with auth, payments, custom domains and real users — plus the creative frontend craft that proves I think in pixels and shaders too, not just systems.

7 projects · disassembled from the core

NavikLab object

NavikLab

A live, monetized SaaS that scores how developers steer AI agents across six cognitive dimensions — auth, Stripe, a cognitive-API moat. YC S26 / a16z applied.

Next.jsSupabaseStripeLive $49/mo
View details CodeLive
Discuria object

Discuria

Daily critical-thinking training — real-world scenarios, neural TTS, ranked mode, and IQ-style scoring. Live and shipping.

Next.jsNeural TTSLive
Open project CodeLive
Smarnh object

Smarnh

AI-powered async group reaction videos — react together, remember forever. Shipped on a custom domain.

ReactAI DirectorLive
Open project CodeLive
ThinkForge object

ThinkForge

A GitHub for thinking — show HOW you think, not just what you know. A cognitive scoring platform.

Next.jsCognitivePlatform
Open project Code
Animated WebGL Library object

Animated WebGL Library

86+ standalone WebGL/Canvas visualizations — zero dependencies, zero build tools. Pure browser craft.

WebGLGLSLCanvasZero-Dep
View details CodeLive
Mole World Dashboard object

Mole World Dashboard

216 Claude sessions became a film-production studio — real-time pipeline monitoring, storyboard viewer, and a voice lab.

Three.jsSupabaseReal-time
View details Code
Podcast Factory object

Podcast Factory

A multi-brand podcast pipeline — script generation, TTS, video rendering, and publishing, end to end.

TTSRenderingPipeline
Open project Code

Experience

13+ years across financial services, retail, and AI platforms.

Senior Consultant — Frontend Lead

FINRA · Toronto / New York

Dec 2021 — Present
  • Redesigned Angular + NgRx workflows that improved analyst productivity by 30% across the regulatory platform.
  • Defined end-to-end testing strategy (Cypress, Playwright, Jasmine), enabling confident bi-weekly releases for a compliance-critical platform.
  • Built AI tooling: coding agent for test generation, design-to-code tool, MCP-based automation for code reviews, PR workflows, and Jira analysis.
  • Prototyped agent swarm for autonomous task execution and a classification model predicting review case complexity.

Senior Frontend Developer

Home Depot · Toronto

Oct 2019 — Nov 2021
  • Architected organization-wide Angular component library adopted across multiple customer-facing and internal apps — modular, tested, accessibility-compliant.

Full Stack Developer

CIBC · Toronto

May 2019 — Oct 2019
  • Built payment platform for international student fee disbursement — Angular 5, Node.js/Express, MongoDB.

Full Stack Developer

TD Bank · Toronto

May 2018 — Feb 2019
  • Built credit card application with instant approval flow — Angular 6, TypeScript, Java/Spring backend.

Full Stack Developer

Nico Info Systems · Hyderabad, India

Aug 2012 — Jul 2016
  • Built enterprise web applications (Java, Spring Boot, Angular) — production-grade UIs with RESTful backends.

Technical Skills

Across the full stack — from pixel-perfect UIs to GPU inference.

Frontend

AngularReactNext.jsTypeScriptNgRx/Redux/RxJSZustandTanStack QueryTailwind v4shadcn/uiThree.js/WebGLStorybook

AI Engineering

OllamaMulti-Provider RoutingRAG PipelinesBM25 + Vector SearchSemantic CachingQLoRA Fine-TuningLLM-as-Judge EvalMulti-Agent OrchestrationLangGraphMCP SDK

Backend & Data

FastAPIHonoNode.js/ExpressJava/SpringPostgreSQLMongoDBRedisQdrantChromaDBDrizzle ORMDocker

Testing & DevOps

CypressPlaywrightVitestJestJasmineGitHub ActionsJenkinsPrometheusGrafana

Certifications

Machine Learning Specialization

Stanford University / DeepLearning.AI (Coursera)

Supervised LearningNeural NetworksDecision TreesUnsupervised LearningRecommendersReinforcement Learning

Let's build something.

Looking for a senior engineer who can lead frontend architecture and build agentic AI product features end to end? Let's talk.

deepchand89k@gmail.com github.com/aptsalt