Abhiram Shaji
Solutions Architect III · Forward Deployed AI Engineering
Montreal, QC, Canada · Open to relocation across British Columbia, remote, hybrid, or on site
236-255-3669 · write4abhiram@gmail.com
At Botpress, I take messy problems that come in from the sales team, design the architecture, and build and ship the solution end to end using our SDK. Over time, I've become the bridge between enterprise client problems and gaps in our SDK. I solve the problem for the client, then turn the solution into something reusable so the next deployment is faster. That sits on two years of professional engineering across Botpress and Deloitte, building enterprise systems from client problem to production. I do my best work when I'm owning the problem with the customer, not coding in isolation. I'm also a heavy user of AI assisted engineering workflows and maintain my own harnesses around models, often using parallel subagents coordinated by a manager agent.
TypeScript · JavaScript · Python · SQL · Node.js · Botpress ADK · LLM agents · RAG · semantic search · structured extraction · evaluation · REST APIs · webhooks · Docker · CI and CD
Solutions Architect III · Forward Deployed
Botpress
Montreal, QC, Canada
- Own 19 enterprise AI deployments across North America and Europe, from scope and architecture to implementation, testing, launch, and handoff.
- Joined as the second person on the delivery team and built the engineering harness the team still ships on as it grew from 2 to 8 people.
- Turn recurring client problems into reusable packages, tools, and product patterns instead of solving the same problem twice.
- Build production agents using tool calling, RAG, semantic search, structured extraction, authentication, long running workflows, and human escalation.
- Built evaluation pipelines that run deployed agents through real channels and grade conversations against expected behavior using LLM judges.
- Ship across webchat, email, WhatsApp, Instagram, support, CRM, and commerce workflows.
Associate Analyst
Deloitte USI Consulting
Bangalore, India
- Built automation for Aflac that categorized and routed insurance claims requiring revision.
- Built internal dashboards, backend services, and Selenium regression automation using .NET, JavaScript, and SQL.
PetSmart · Yamaha Motor Mexico · iolo Technologies · HEAG · London and Partners · IonOptix · YWCA Calgary · CFL Flooring
Postgraduate Diploma, Digital Design and Development
North Island College
BC, Canada
Bachelor of Engineering, Computer Science
Bangalore City University
Bangalore, India
Dear hiring team,
I build production AI agents. At Botpress I take the spec, own the architecture, write the TypeScript that ships, then document it and hand it to the support team.
In practice that means agents that qualify leads and hand them to a CRM, retention agents that read a real account before they answer, compliance agents that answer only from source documents and cite them, and support agents that authenticate a user before touching a refund. They run on whatever channel the client already uses, webchat, email, WhatsApp, Instagram, Facebook, and some run inside our own ticketing platform, Botpress Desk, where the agent answers the ticket, leaves internal notes for staff, and routes it to the right team. When a question sits outside what the agent can prove, it escalates to a human rather than guessing.
The work that repeats gets promoted into the product. I turned the commerce delivery into versioned packages inside our agent framework, a headless core carrying the store contract, catalog search and money helpers, and the tool factories for product search and product details, with a separate binding layer for the chat UI. The core stays free of React and webchat on purpose, because that boundary is what makes it reusable rather than one client shaped. Its card views install through a small CLI I wrote that keeps a checksum manifest, so a bot can tell when its copy has drifted from the package and a collision is refused instead of silently overwritten.
I also set the structure the team codes to. Shippable code and throwaway code are separated by folder, so promoting a prototype is a file move rather than a naming convention nobody follows. Bots are atomic, one tool, one table, one handler per file. Branches follow a validated pattern, new projects come out of a scaffolder, and every bot in the fleet has a generated metadata record, so channels, integrations and patterns are readable without opening the code. Debugging production runs through an extended CLI I maintain for logs, conversations, exports, and comparing one environment against another.
On top of that I built the harness the team delivers on. It takes a client spec, plans the file layout, builds it with parallel agents, then fails the build on tools nothing calls, modes nothing reaches, and placeholder values left behind. Most agent bugs I have shipped were things that were never wired up at all.
Evaluation is the part I care about most. Tests drive the deployed agent over its real channel and an AI judge scores the transcript against what should have happened, so a fluent answer with no source fails. In production the agent carries its own judge, scoring every finished conversation against a rubric, and a fallback row is written when the judge call fails, because losing coverage quietly is worse than a bad score.
The same habit shows up in my own setup, which is Neovim and Ghostty with Claude Code alongside. Most of it is a Lua control panel I wrote for the parts of the day that are not writing code, picking a repo, opening a branch or a pull request in its own worktree, reading tickets, keeping the editor and terminal themes in step. It is a few thousand lines to save keystrokes, which I accept is a personality trait.
What I want next is the same work with more ownership of the product around it. I am looking for forward deployed, solutions architecture, or applied AI engineering work, remote or anywhere in British Columbia.
Thank you for reading.
Abhiram Shaji