AI inside real products, under real constraints

Each user's browser creates a key pair and protects it with a PIN before it leaves the device, so the platform stores only encrypted keys. After payment, the agency releases the transcript key to the attorney without the server reading it.

Ruled outServer held keys, which would let the platform read transcripts.

What it gave the clientAdmins cannot open delivered transcripts, while payment, release and audit still run through the platform.

Malware, content, duplicate and format checks run on the seller's own device, and only manifests, metadata, hashes and licenses reach the platform, even for files up to 10 GB.

Ruled outUploading datasets to a central server for checks.

What it gave the clientNo dataset file ever passes through the platform.

Runtime secrets moved to the client's AWS Secrets Manager per environment, read through scoped roles with no write access for us. Production access became just in time, for a short named list.

Ruled outHolding production keys in our own tooling.

What it gave the clientSales never stopped, and every control now sits in accounts the client owns and can show a reviewer.

Contracts never leave the client's restricted AWS account. One RAG pipeline runs there on Lambda for every contract layout, and returns NULL when a term is absent.

Ruled outParsing rules written for each contract layout, which broke on the next format.

What it gave the clientLive on Lambda in the client's own AWS account, approved after the first feedback round.

The client's backend APIs became a tightly defined tool registry, called under each user's own Cognito identity, so the model acts only through real endpoints and existing access rules.

Ruled outA separate AI backend with its own copy of the data and its own permission model.

What it gave the clientLive for every role, through the same APIs and logins the product already used.

A confident score built on another company's evidence is a failure nothing downstream catches, so companies match only on domain, business number or exact name, and the engine reads the platform through a read only role.

Ruled outFuzzy name matching and auto matching on name alone.

What it gave the clientLive on the client's production domain, with a scoring method the client owns.

The existing application stays the system of record, reached only through its APIs. AI drafts sit next to the human entered deal, and a person's final version wins.

Ruled outWriting AI output straight into the live deal record.

What it gave the clientIssuers no longer retype documents into deal forms, and every regulated decision stays with a person.

Each phone provider reports calls its own way, so typed models check every raw event, and a cloud function turns it into the product's own call records inside its existing API and permission model.

Ruled outWriting raw events straight into the product's records.

What it gave the clientA tested path from each provider's raw events into the product's own call records.

Before rollout, a layered test set drawn from the client's own business rules ran against ten approved reporting views, and every figure had to match the existing reports.

Ruled outLetting the model write and run any query it liked.

What it gave the clientA clear verdict for the client: reliable on direct questions over its views, with two limits and a known fix.

Agents grade each join against the data catalog, and a guessed join pauses the build for a person. Every query passes an EXPLAIN check before any sample run.

Ruled outLetting the model infer joins from column names.

What it gave the clientAn evals harness gives the founder a measured baseline for the build agent, not a demo.

LanguageTool finds the errors, the model is asked only about ambiguous spans, and only those span edits are applied, so approved names, numbers and attributions never change.

Ruled outLetting the model regenerate the whole passage and hoping the facts survive.

What it gave the clientApproved names, numbers and attributions stay untouched, in English and French.

The copilot runs on model APIs inside the customer's own environment, with access by role, and explains the SQL behind each answer so a wrong number is caught before anyone acts on it.

Ruled outA chat answer with no way to check where the number came from.

What it gave the clientIn use at two companies before it became a multi tenant product.

Each document's permissions travel with it into every answer: results are filtered by the user's clearance before ranking, and when access cannot be confirmed, nothing is returned.

Ruled outFiltering after generation, where a restricted passage could already have shaped the answer.

What it gave the clientIn the client's own tests, the system stopped inventing names when documents lacked the answer.

A single practice intake product became multi clinic, with super admin, clinic admin and provider roles, each clinic isolated, and patient identifiers removed before any call to the model provider.

Ruled outSending full intake records, names included, to the model.

What it gave the clientNow a commercial product clinics sign up for, with the code, billing and secrets owned by the founders.

Access moved from cookies to JWT on every API route, and self hosted Supabase with row level security gave one consistent check, ready to install on a customer's network.

Ruled outRelying on application code alone for isolation.

What it gave the clientIt can now run hosted, with a customer's own keys or inside a customer's network.

The assistant answers only from the documents uploaded for the deal in view, keeps no memory between sessions and gives recommendations only.

Ruled outA general assistant over all deals, where answers could mix one company's facts with another's.

What it gave the clientInvestors get answers drawn only from each deal's own evidence, with every decision left to people.

A desktop app records any meeting tool through a signed virtual microphone. Recording can never break the agent's live audio, and an upload queue recovers when the laptop goes offline.

Ruled outA capture path that could silence the agent mid call.

What it gave the clientIn production with the client's agents, and it kept running while users moved to a new platform.

We recovered the legacy diligence database a former lead had left without documentation and moved it with ETL scripts, with AI suggesting mappings from the old schema, while the old portal stayed up.

Ruled outRekeying manager records or switching the old portal off first.

What it gave the clientAnalysts, investors and sub-clients now work from one platform, with restricted data inside the roles that own it.

Tests on a demo account showed what the gateway's hosted form could not do before release, a small token service on its own whitelisted server kept card data off our backend, and only production moved.

Ruled outChanging the card flow so card data passed through our backend.

What it gave the clientOrganizations kept the same screens and data, and card data never reached the backend.

We diffed the fork against its base and the latest release, moved every custom feature across inside the client's own cloud accounts, then added a gateway for many models.

Ruled outAdding hosting and new features on the old fork and upgrading later.

What it gave the clientCustom features on a current base it can keep upgrading, with new AI providers added without code changes.

Answers come from each dealer's feed, synced into Elasticsearch every hour, so the assistant never invents stock or prices, and towing fit is calculated in a tool, not by the model.

Ruled outAsking the model whether a truck can tow a trailer.

What it gave the clientLive on dealer websites, with stale prices fixed by re-indexing only the units that changed.

Match scores come only from the founders' Snowflake engine, with the configuration version stamped on every run. The AI drafts explanations and outreach from the score breakdown and source records.

Ruled outLetting a model adjust match scores.

What it gave the clientOrganizers go from an event to scored, explained sponsors and a grounded outreach draft, every credit metered.

A deterministic scoring engine on structured financial data picks the approach, and the model writes the lead profile and outreach draft from a fixed payload.

Ruled outLetting the model decide which angle to take with a prospect.

What it gave the clientLive and selling, with exclusive leads and a reasoned approach instead of shared lists.

With no labeled ground truth, regex rules give a fast, explainable anchor and the model's suggestions are merged through calibrated weights, around the founder's scoring logic left intact.

Ruled outLetting the language model decide the classification.

What it gave the clientClassification became consistent, with every tier working end to end.