The demo that never shipped
A notebook is not a product. Shipping needs the interface, the backend, the deployment and the tests — not just a prompt.
Research first. Revenue next.
North Summyn tests AI agents before anything gets built. In a two-week Base Camp Sprint we attack your agent in staging, measure what it costs to run, and hand you every failing case, the fix, and a fixed price to close it. Then we build what the numbers justify.
Book aBase Camp
Sprint
Two weeks · Fixed price · Staging only · The report is yours either way
Scroll to begin the climb
IThe fog
Most AI projects never leave the demo. The prototype impresses, then production arrives — and the context doesn't fit, the answers can't be trusted, and the bill keeps climbing.
The limit that lived in the prompt
A rule written in the system prompt is a suggestion, not a control. In our tests, limits that lived only in the prompt were broken in 21% of attacks. Limits enforced in the tools were broken in none.
The bill that kept growing
Every unnecessary token and every avoidable LLM call adds up. Cost has to be engineered in from day one.
IIThe lab
We build our own tools before we build yours.
Research · Agent security
Agent security lab
We attack agents the way attackers will, and score every run by machine.
Read the studies →- Study 1: 27 attacks, 3 models, 480 valid runs. Limits in the prompt broke 21% of the time; limits in the tools, 0%.
- Study 2: one hostile third-party tool server poisoned descriptions and results with no jailbreak and no hostile user.
- Six attack classes from direct injection to tool shadowing, each with a machine-checkable success test.
- Every study publishes its limits. The hardened 0% is by construction; the baseline is the finding.
Python · tool-level guards · canary strings · Wilson intervals · open-weight and frontier models
Open source · Codebase intelligence
karst
Know what your change breaks — without your code leaving your machine.
github.com/Moin105/karst →202tests, up from 125 across two releases
6languages parsed with tree-sitter
0telemetry — indexing runs fully local
- AST-aware chunking across Python, JS, TS, Go, Rust and Java
- Hybrid retrieval: a call/import graph combined with vector search, returning scoped context with exact file:line citations
- MCP server for Claude Desktop, Cursor and any MCP client
- Local ONNX embeddings for air-gapped and regulated environments; incremental re-indexing by content hash
Python · tree-sitter · Qdrant · NetworkX · MCP · FastEmbed/ONNX · SQLite
Autonomous agents · B2B research
AgencyOS
An agent that plans its own research, then executes it.
Private — walkthrough on request →- A planner agent composes each day's research plan from goals and accumulated memory — outcomes feed back, so later plans improve
- Discovery agents source companies from public data with a polite crawler that respects rate limits and robots.txt
- A deterministic prefilter removes low-signal leads before any LLM call, cutting the cost per qualified lead
- An AI analyst produces structured, explainable weighted scoring for every surviving candidate
FastAPI · async SQLAlchemy · PostgreSQL · Alembic · Claude / Gemini · APScheduler · Docker · pytest
Open source · LLM cost engineering
TokenTier
Send every task to the cheapest model that can handle it.
github.com/TokenTier →- Cost-aware model routing for Claude Code: starts on Haiku and escalates to Sonnet, then Opus, only when a concrete check fails
- Logs every task locally and shows cost and savings against an all-Opus baseline in a local dashboard
- Merges safely into existing settings, backs up every file it touches, and uninstalls exactly what it changed
- Standard-library Python with no external dependencies; runs on macOS, Linux and Windows
Python · Claude Code hooks · subagents · local dashboard
IIIThe route
Four camps on the way up.
-
Camp I
Agent Security & Cost Engineering
Attacks written for your tools and run in staging, limits moved from the prompt into the code, approvals taken out of band, and model routing, context trimming and prefiltering that bring the bill down.
↗ -
Camp II
AI Agents & Multi-agent Systems
Agents that plan, use tools and remember — with structured outputs, evals and human approval where it matters.
↗ -
Camp III
RAG, Retrieval & MCP Servers
Chunking, embeddings, hybrid search and reranking — exposed to your AI tools through MCP, with citations you can verify.
↗ -
Camp IV
Full-stack AI Products
The whole product around the model — React and Next.js interfaces, FastAPI and NestJS backends, AWS and Docker deployments.
↗
✦Selected work
Sixteen years of shipping, combined.
Work our co-founders led before North Summyn, at the companies named.
180+APIs integrated in two weeks
30+developers led
16years of combined experience
Inresi
A real-estate experience rebuilt around immersive 360-degree environments, so buyers can walk a property from the browser.
Next.js · React · Redux · Three.js · AWS S3vortic.art
A virtual-reality and digital exhibition ecosystem with live events and VR support for galleries and artists.
vortic.art ↗Vortic CMS
The content management system galleries use to build and run their exhibitions, with payments built in.
React · Redux · Tailwind · Stripe · JestIn-house financial platform
Designed and built a platform for managing and tracking company finances — 180+ APIs integrated within a two-week window, a Next.js REST layer with Swagger docs and Jest tests, plus a custom CRM on AWS. Led a team of 30+ developers.
React · Redux · Next.js · Tailwind · Jest · Swagger · AWSInternal developer marketplace
An internal enterprise marketplace serving 5,000+ developers across six integrated marketplaces.
React · TypeScriptArcade platform in two weeks
A full-stack arcade platform with a canvas game engine, Firebase authentication and a realtime leaderboard — delivered in two weeks.
Canvas · Firebase · RealtimeSignature PWA
An offline-capable progressive web app that turns drawn signatures into verifiable PDFs.
signature-pwa.vercel.app ↗A 500-developer codebase, rebuilt
Re-architected a large legacy Vanilla JS and Laravel codebase originally written by 500+ developers — 450 focused hours on a one-month timeline, improving performance, scalability and maintainability.
React · Next.js · TypeScript · TailwindAlso built: a custom Ethereum network and its block explorer (Fornax), a smart-contract marketplace (DappsLab), and Web3 wallet integrations (SmartWhales).
IVThe ascent
0000m
01 — Base camp · 2 weeks · fixed price
Base Camp Sprint
We attack your agent in staging with tests written for your tools, review what it costs to run, and agree how success will be measured. You get every failing case, the fix, a fixed quote, and a report that is yours whether or not you climb further.
02 — First ridge
Prototype
A working slice on your real data, fast. If it doesn't earn its place, we stop before you've spent much. The sprint fee comes off the build price if you sign within 30 days.
03 — High camp
Build
Production engineering: evals, monitoring, cost controls, tests and deployment — with humans in the loop for anything sensitive.
04 — Summit
Climb
We keep improving against the numbers that matter — quality, latency and cost per task.
VThe studio
Led by two founders who ship.
Backed by a small delivery team we bring in per project.
Moin Latif
Six years of production experience — from pixel-perfect frontends to agent orchestration, hybrid retrieval and LLM cost engineering. Has led a team of 30+ developers, re-architected codebases at scale, and taught blockchain and smart-contract development as a certified instructor under Pakistan's National Skills Development Program (NAVTTC × NUML).
- AI / LLM
- MCP · RAG · hybrid search · agents & tool calling · evals · Qdrant · pgvector
- Stack
- TypeScript · Python · React · Next.js · FastAPI · NestJS · AWS · Docker
Qasim Raheem Khokhar
A decade of building across the stack — senior engineering at Superbright and Vortic, blockchain engineering at Fornax and DappsLab, and earlier lead roles at Helfer and BCDApps. Team lead, project manager and blockchain full-stack developer, from immersive 3D and VR web apps to custom Ethereum networks and smart contracts.
- Blockchain
- Solidity · Web3 · Ethereum · ERC standards · Geth · Truffle · Ganache
- Stack
- React · Next.js · Node.js · NestJS · Golang · PostgreSQL · MongoDB · AWS · Docker
VIThe summit
01
Shipped over demoed.
A system counts when it runs in production — with tests, monitoring and a deploy you can trust.
02
Evidence over opinion.
Every model, prompt and retrieval choice is judged by evals and real numbers, not vibes.
03
Your data stays yours.
Local-first where it matters, human approval for anything sensitive, and zero surprise telemetry.
✦North star
Find your north.
Reach your summit.
Tell us what your agent can do today: which tools, which actions, where the limits live. We'll tell you honestly what we'd test, what it would cost to run, and whether a sprint is worth your two weeks.