Skip to content
JP Levi AI
(929) 356-4644

What we
actually do

08 practices

Eight practices. Most projects use two or three of them, and the honest answer to which ones you need usually comes out of a short call rather than a page like this.

  • Retrieval

    Find the right context across your data.

  • Reasoning

    Infer, synthesize, and generate with precision.

  • Deployment

    Ship, monitor, and operate with confidence.

Private knowledge systems

01 / RAG · GraphRAG

A search box that answers questions about your own documents, and shows you which page it took the answer from.

You would want this if

  • Staff keep asking each other the same questions and digging through folders to answer them
  • You have years of contracts, reports or manuals and nobody can find anything in them
  • You tried a public chatbot and it confidently invented things about your business

What you get

  • Ingestion, parsing, and chunking pipelines
  • Vector and hybrid search tuned on your queries
  • GraphRAG over entities and relationships
  • Evaluation sets that prove quality before launch
Technical detail

Answers drawn from your own documents, with citations back to the source. Graph retrieval when the relationships between entities matter more than the passages.

Retrieval
vector · hybrid · graph
Stores
pgvector · Neo4j
Evaluation
golden sets · regression gates

MCP servers and connectors

02 / Model Context Protocol

A standard plug that lets AI assistants use the systems you already have, safely, without handing your data over to them.

You would want this if

  • You want staff to ask an assistant about live data sitting in your CRM or database
  • You keep paying to rebuild the same integration every time a new tool appears
  • You need a record of exactly what the assistant was allowed to see and do

What you get

  • Tool, resource, and prompt definitions
  • Auth, scoping, and audit trails
  • Local stdio and remote HTTP transports
  • Client wiring for the apps your team already uses
Technical detail

Your systems exposed as tools any model can call. Built once against the protocol, then usable from Claude, an IDE, or your own application.

Transport
stdio · http · sse
Auth
oauth2 · api key
Clients
Claude · IDEs · your apps

Agents and automation

03 / Tool use · workflows

Software that carries out multi-step jobs on its own, and stops to ask a person whenever being wrong would be expensive.

You would want this if

  • Someone on your team spends hours moving information between systems
  • A process is well understood but tedious, repetitive and easy to get wrong
  • Work needs to happen overnight, on a schedule, or the moment something arrives

What you get

  • Workflow design with explicit failure paths
  • Tool use, retries, and budget ceilings
  • Human-in-the-loop review gates
  • Queues, schedulers, and background jobs
Technical detail

Multi-step work that runs without a person driving it, with guardrails and human review at the points where being wrong is expensive.

Runtime
queues · cron · webhooks
Controls
budgets · timeouts · review
Tracing
per-step logs and replay

Machine learning

04 / Forecasting · scoring

Predictions built from your own history: what is likely to sell, who is likely to leave, which record looks wrong.

You would want this if

  • You forecast demand, stock or staffing with a spreadsheet and experience
  • You want things ranked or scored automatically instead of by hand
  • You need unusual or suspicious activity flagged before a person sees it

What you get

  • Feature engineering on your historical data
  • Training, validation, and honest error analysis
  • Batch and real-time inference services
  • Drift monitoring once it is live
Technical detail

Trained models for the problems where a prompt is the wrong tool: forecasting, scoring, ranking, classification, anomaly detection.

Stack
PyTorch · scikit-learn
Serving
batch · realtime API
Monitoring
drift · data quality
Consult with an expert ↗or call (929) 356-4644

Web and product engineering

05 / Full stack

The application itself: what people see, what happens when they click, and everything running behind it.

You would want this if

  • You need the thing built, not just advice about building it
  • Your current site cannot do what the business has grown into needing
  • You want the AI parts and the app around them built by the same people

What you get

  • Next.js and TypeScript front ends
  • Auth, billing, roles, and admin tooling
  • Design systems that survive a second developer
  • Static or server-rendered, chosen per job
Technical detail

Complete applications, front to back. The AI parts built in from the start rather than bolted on to something that cannot carry them.

Front end
Next.js · React · TS
Back end
Node · Python · FastAPI
Delivery
CI, preview builds, rollback

APIs and data engineering

06 / Pipelines · integration

Getting your systems to talk to each other, and keeping the numbers right while they do.

You would want this if

  • Two systems hold the same information and quietly disagree
  • Someone exports a spreadsheet by hand every week so a report can exist
  • You are moving off spreadsheets, or off software you have outgrown

What you get

  • REST and typed API design
  • ETL and sync between third-party systems
  • Schema design, migration, and cleanup
  • Backfills that can be run twice safely
Technical detail

The plumbing under everything else: moving data between systems that were never designed to talk, and keeping it correct.

Databases
Postgres · SQLite · Redis
Integration
webhooks · queues · cron
Quality
idempotent, replayable jobs

Managed hosting and operations

07 / VPS · DevOps

Hosting detail ↗

We run the servers your software lives on: keeping it online, backed up, patched and watched.

You would want this if

  • Nobody in-house owns the servers, and it shows when something breaks
  • You have backups but nobody has ever tried restoring one
  • You want one place to call when something breaks, instead of three vendors pointing at each other

What you get

  • Provisioning, hardening, and TLS
  • Deploy pipelines with a tested rollback
  • Restore-tested backups, not just backups
  • Uptime and error alerting that reaches a human
Technical detail

Servers we provision, harden, monitor, and keep patched, so the thing that was built stays running after the build ends.

Platform
Linux · Docker · nginx
Delivery
push to deploy · rollback
Watch
uptime · disk · errors

Internal tools for small teams

08 / SMB systems

Small pieces of software shaped around how your team already works, instead of bending the team around someone else's product.

You would want this if

  • Your process lives in one person's head and a spreadsheet only they understand
  • Off-the-shelf software does most of it and you work around the rest
  • You need reporting your team will actually open and use

What you get

  • Automation across tools you already pay for
  • Dashboards and reporting people actually open
  • Data cleanup and migration off spreadsheets
  • Documentation and handover
Technical detail

Automation and internal software for companies without an engineering department, built so the team can operate it without us.

Shape
internal apps · automations
Handover
runbooks · training
Lock-in
none, the code is yours

Not sure which of these you need?

Most people are not, and that is fine. Describe the problem in plain words and we will tell you which of these applies, or that none of them do and what to do instead.

Start a conversationor call (929) 356-4644