Hybrid retrieval
Lexical · semantic · AST · graph
Suhail Shah
I'm Suhail Shah, an applied AI and full-stack software engineer who builds production AI systems, developer tools, and the backends that keep them reliable.
A multi-provider AI coding agent that grounds work in AST-aware repository context and trusted knowledge, cites its sources, and verifies changes with approved tools on the developer's machine.
Production systems built for real business operations, combining product thinking, scalable architecture, and end-to-end engineering.
Senior Full Stack Engineer
Led the architecture and full-stack development of an enterprise workflow and approval platform, turning complex approval processes into reusable workflow and form templates. The platform became a shared foundation for request intake, role-based collaboration, approvals, documents, audit history, dashboards, PDF exports, and admin-managed configuration. It helped teams consolidate scattered processes and reduce reliance on separate tools or ad-hoc systems without hard-coding each new workflow.
Senior Full Stack Engineer
Built an event security operations platform that unified event setup, shift planning, guard records, attendance verification, live operations monitoring, incident reporting, patrol tracking, audited overrides, and invoice discrepancy review. The platform gave operations and commercial teams a reliable source of truth for staffing coverage, verified hours, no-shows, exceptions, and supporting evidence, while integrating with a separate Android guard app through shared API contracts.
Full Stack Engineer
Built a single-tenant internal operations platform for a wholesale distribution business that had outgrown spreadsheets. The system replaced separate stock and order workbooks with one database-backed workflow covering products, customers, stock movements, sales orders, purchase orders, warehouse picking, stocktakes, dashboards, reports, audit history, and accountant-ready exports.
Software Architect
Architected and built a schema-driven form engine for a large enterprise client whose systems centered on multi-step form journeys. The engine let delivery squads declare journeys as JSON, render them through a shared React library, validate them again on the Node.js backend, version schemas centrally, and release new field capabilities through a private npm package workflow.
Full Stack Engineer
Worked across an investor e-services platform supporting shareholder submissions, IPO-related flows, internal back-office operations, and corporate website maintenance. Built and extended form modules, maintained PHP features, wrote MySQL queries and reports, and handled front-end work with HTML, CSS, JavaScript, and jQuery.
Featured product case study
A production AI coding agent that keeps code execution local while a multi-provider backend plans, validates, and verifies every change. It combines AST-aware repository context with trusted knowledge, cited evidence, document intelligence, and real project checks.
Lexical · semantic · AST · graph
Evidence stays attached to answers and plans
Format · compile · lint · test · repair
Files and commands stay on the developer’s machine
The product’s key design choice is simple: AI can reason about a codebase, but the user’s machine stays in control of its files.
A developer works in the fourpoket CLI rather than handing a repository to a remote agent.
The CLI maps code with AST and graph relationships while hybrid retrieval finds relevant repository context and trusted knowledge.
The backend combines cited sources with validated typed actions, shows the plan, and requests approval for state-changing work.
The CLI edits locally, runs approved project checks, supports bounded repair, and reports the diff, evidence, usage, cost, and outcome.
The difficult part was not adding a chat box. It was making AI-driven changes reliable, inspectable, and safe to operate.
The backend decides what to do. The CLI is the only part of the system that reads and writes the developer’s files.
Named blocks keep provider output reliable while a canonical typed action layer gives the rest of the system one validated contract.
AST-aware targeting, hybrid retrieval, incremental mapping, and pruning keep sessions focused while usage and evidence stay visible.
Versioned knowledge collections let fourpoket work with current technologies, private engineering material, and complex documents.
A generated patch is not considered complete until approved local tools have checked it and the results are visible.
TypeScript owns product behavior and authorization; Python provides isolated AI computation behind versioned contracts.
Each system owns a distinct responsibility while sharing versioned contracts, session state, usage, account, and product rules.
Express control plane for agent orchestration, validation, accounts, billing, knowledge ingestion, hybrid retrieval, citations, evals, and business logic.
RailwayInk terminal UI for AST and graph mapping, knowledge commands, local reads and writes, approvals, configured tool execution, and evidence-rich session receipts.
npmProduct site, account flows, dashboard, knowledge collection management, document review, session citations, usage history, and Stripe balance purchasing.
VercelInternal diagnostics for sessions, users, ingestion jobs, retrieval traces, model workers, evals, pricing safeguards, revenue, and support operations.
Local / DockerPrivate Python service for document intelligence, embeddings, reranking, model inference, and reproducible evaluation workloads.
Private / DockerSelected stack
TypeScript · Node.js · Express · React · Next.js · Ink · Python · FastAPI · SQLite · PostgreSQL · pgvector · Cloudflare R2 · Zod · Tree-sitter · PyTorch · Stripe · Docker
Personal AI systems work, built in public with readable source.
Public source, actively developed
A local learning platform where an LLM acts as the teacher and a typed Python backend owns memory, orchestration, and context engineering. The public source spans the applied AI stack: provider abstraction over two model transports, pgvector retrieval, a versioned eval framework with regression reporting, OpenTelemetry tracing, and an approval-gated agent layer.
Hey, I'm Suhail Shah. I build applied AI and full-stack products where the models, frontend, backend, data layer, and product behavior all have to work together in production.
Here's the bet I'm making with my career. Every powerful tool has always charged an entry fee. Blender, Photoshop, a serious spreadsheet, each one takes months before it gives anything back. AI collapses that fee into a conversation. You say what you want and the tool meets you there. Wrapping hard software in plain language is the biggest shift in how people use computers since the GUI, and it gets won or lost at the application layer.
The chat box is the easy part to see. The hard part is everything behind it. The system has to understand intent, find the right context, use the right tools, check its own work, recover when it is wrong, and stay cheap enough to run. That is why full-stack matters here. The magic keeps moving between interface, backend, data, permissions, prompts, evals, and deployment. A good AI product is not just a model with a nicer input. It is a full piece of software wrapped around a new way of asking.
Suhail is often associated with Canopus, one of the brightest stars in the night sky. Across history, Canopus has been used as a navigational star, a fixed point for finding direction across long distances.
I like that connection because it matches the kind of engineering I care about: making complex systems easier to navigate, turning ambiguity into direction, and building tools people can trust when the path is not obvious yet.
Interested in working together? Let's connect.
Open to AI Engineer roles · Remote