updown move enter open esc close matheesan.ca / index
    2018 first enterprise system Avaya · real-time comms
    2019 first production ML CGI · document intelligence
    2021 first LLM agent voice + tool use
    2026 marketplace platform WayTrade · biofuel

    The agents are the visible artifact. The job was still software engineering: API reviews, state machines, integrations, infra, tests, release trains, incident response — and choosing the tooling that made each loop shorter. see the tooling timeline

    Nov 2026 — Present Remote

    Senior Software Engineer

    most recent

    WayTrade

    Senior software engineer on a small team building WayTrade’s AI-powered marketplace for biofuels, feedstocks and critical materials — the commercial, operational and compliance work behind a physical trade.

    • Building the marketplace platform end to end: trade records, offers, logistics, inventory, documents and finance in one system.
    • Working across the full trade lifecycle, where a delivery change cascades into inventory, production, cash and compliance.
    • Applying AI where it removes friction — matching supply and demand, surfacing exceptions and recommending the next action — with the trade record as the source of truth.
    marketplace platformsproduct engineeringAI workflowstrade operationscompliance
    Oct 2021 — Oct 2026 Remote · Toronto

    Senior Software Engineer

    Replicant

    Forward-deployed AI engineer: enterprise LLM voice agents, model routing, knowledge/RAG layers, evals and agentic tooling — from API review through launch and hypercare.

    • Designed and built the org-wide agentic workbench solo: one engine connecting multiple data silos and workstreams through plugins and MCPs, with DeepSeek/Claude-powered overnight planning, scoping, building and JIRA organization behind gates.
    • Took enterprise LLM voice agents from an empty repo to production: state machines, tool calls, auth, fallbacks and hypercare.
    • Built the platform layer around them: provider routing and fallbacks, knowledge bases and RAG, eval harnesses, guardrails and observability.
    • Worked across frontier and open-weight providers; treated model and prompt upgrades as releases with regression evals.
    • Caught launch-blocking auth, wrong-record and first-time-user defects before they reached callers; turned them into acceptance tests and release gates.
    • Built agentic tooling — AI coding skills, MCP servers, code-enforced reviewer gates — and drove Copilot, Cursor and Claude Code adoption across engineering.
    LLM systemsmodel routingRAGvector searchTypeScriptPythonGCPMCPClaude Code
    2019 — 2021 Toronto

    Software Engineer — Emerging Tech

    CGI

    Enterprise AI/ML and MLOps in the pre-LLM era. Built document-intelligence and model-delivery prototypes when both were still bespoke.

    PythonOCR / NLPMLOpsAzureDockerCI/CD
    2018 Toronto

    Software Engineering Intern

    Avaya

    Enterprise communications infrastructure. Learned to debug real-time systems where a second is a long time.

    SIP / VoIPnetworkingtest harnessesLinux
    2015 — 2019 Toronto

    BEng, Computer Engineering

    Ryerson University → TMU

    Computer engineering with a final-year ML project: Twitter data into interaction graphs into community structure.

    PythonNetworkXNLPgraph analysis

    The agentic workbench.

    The most important project I built at Replicant. One engine that connected multiple data silos and streams of work through plugins and MCPs, then let DeepSeek or Claude run the night shift: planning, scoping, building and organizing JIRA behind gates. Built solo, with my own harnesses.

    solo build org-wide plugins + MCP DeepSeek / Claude overnight loops gated execution
    read the case file

    Before the boom.

    The tools were different; the job was the same: understand the system, build the pipeline, measure the outcome. These are the projects that made the later AI work feel familiar.

    2018 Avaya

    Real-time comms test harnesses

    Built tooling to trace SIP/VoIP call flows and reproduce state across systems that could not be restarted casually.

    Made latency and state visible.

    SIP / VoIPLinuxtest harnesses
    2019 Ryerson / TMU

    Graph communities from Twitter

    Ingested tweets, built interaction graphs and ran community detection on sparse, noisy data.

    First time a model found structure I hadn’t put there.

    PythonNetworkXNLP
    2019 CGI

    Document intelligence before LLMs

    OCR + classical NLP + rules to extract structure from messy documents. The model was a small part; the pipeline was the product.

    Taught me to make the boring path reliable.

    OCRNLPpipelines
    2020 CGI

    The MLOps runway

    Containerised model serving, CI/CD, monitoring and rollback for models that had lived in notebooks.

    Made models deployable, not just demonstrable.

    DockerCI/CDmonitoring

    Under the agents.

    The layer that doesn’t demo: model routing, knowledge bases, RAG and evals. It is the part that makes an agent safe to change after launch — and the part I reach for first.

    01

    Providers

    Frontier and open-weight models behind one interface. Model upgrades are releases, not config changes.

    OpenAIAnthropicGoogleopen-weight
    02

    Routers

    Fallback, retries, cost/latency policy and per-tenant routing. The model is a dependency; the router is the architecture.

    fallbacksretriesbudgetspolicy
    03

    Knowledge bases

    Ingestion, chunking, embeddings, vector search and citations. Retrieval quality is its own eval problem.

    embeddingspgvectorPineconeWeaviate
    04

    RAG

    Retrieval orchestration, context assembly and grounded answers, measured against must-answer and must-not-answer sets.

    retrievalcontextcitationsevals
    05

    Agents

    State machines, tool calls, MCP and guardrails. Deterministic where correctness matters; generative where judgment helps.

    state machinestool useMCPguardrails
    06

    Evals

    Scenario harnesses, must / must-not assertions, replay and production traces — the thing that makes the rest safe to change.

    must / must-notreplaysilent-greentraces

    the model is a dependency the system is the product

    // also built

    • An AI skill that generates every new agent’s API-integration contract from an OpenAPI spec, distributed to every agent repo.
    • A trace-to-fix workbench that turns failing production traces into reviewed draft fixes — with push / merge / deploy refused in code before a model is asked.
    • An unattended outbound dialer on Cloud Run Jobs with a generation-precondition lease, kill switch and Terraform-managed infrastructure.
    • An incident postmortem that corrected an outage window by about an hour, because the worst phase ran under an alert mute.