Four lines of work, deliberately narrow inside each: Data + AI monitoring and alerting, AI agents, applications and the systems around them, and niche products. It's the same kit our own products run on; anything outside it, we subcontract or refer.
Senior pair-hours for teams that already have one. Source selection, scoring logic, prompt and eval design, agent architecture, or an honest read on a build plan — the second opinion before you commit to a quarter of work.
A focused dive through the data you want to watch and how you're watching it. You get a written report with prioritised findings — which sources to trust, what you're missing, and where the false positives are coming from.
A pass over how your LLM reads and scores the data — prompts, relevance logic, and the evals (or lack of them) behind it. We tell you where it's hallucinating, where it's miscalibrated, and what to fix first.
For founders & operators sizing the next quarter of work. We'll read your brief, ask the obvious follow-ups, and come back with a one-page plan you can take to your team — or to a different vendor entirely.
Scoped deliverables across the four lines — a monitoring pipeline, an agent, a whole application, a niche product. Fixed scope, fixed price, weekly demos, written handoff at the end. Where a service below shows no starting figure, the number depends enough on scope that quoting one here would be guesswork; the brief gets a real one within two days.
The engine behind our own products. We watch a source — a public register, an API, your own database — detect what changed, score it, and deliver alerts your team will actually read. By daily digest or in realtime.
LLMs pointed at your data: plain-English summaries, relevance scores, and structured extraction — grounded in the source, with evals you can actually run. Provider-agnostic across OpenAI, Claude, and Gemini.
The unglamorous plumbing that makes the data — and the alerts on top of it — trustworthy. We ingest from feeds, APIs, and databases, then dedupe, entity-match, and normalize so nothing is double-counted or missed.
An agent that does the work rather than describing it — reading your systems, calling real tools, and stopping where it should. Built so failure is loud and cheap: every run leaves a trace, every tool call has a boundary, and a human stays in the loop wherever the cost of being wrong is high.
A whole application, not a prototype that needs a second project to become one. Schema, screens, auth, billing, background jobs, and deploys — bring the idea and the domain knowledge, and there's something on staging in the first week.
For when the software already exists and the problem is that none of it talks. We stand up the infrastructure, wire the third-party pieces together, and build the admin and reporting layer that tells you whether any of it is actually working.
A product for one narrow market, built the way we build our own. Watchlists or workflows, accounts, billing, and the source-linked pages that earn the traffic — on staging in week one, launch-ready at handoff, and yours.
We're explicit about this because it's the fastest way to save everyone a week. If your project is on this list, we know good people; ask and we'll make an intro.
Every project on the same cadence. You'll always know what's next, what we're blocked on, and what's shipping this Friday.
An hour on a call, then a day in your repo and your docs. You get back a one-page plan with a fixed quote and a start date.
A real, working slice — often a first scored record or a sample alert against your own data. A visible stake in the ground that sets the cadence for everything after.
Every Friday: a 15-minute recorded walkthrough of what shipped, what's in review, and what's next. No status decks, no theatre.
Runbook, test suite, and an architecture note you can hand to your next hire. Thirty days of follow-up included, whether you need it or not.
If yours isn't here, send it via the form on the homepage. We answer most questions in plain text the same business day.
Most engagements land between four and ten weeks. A focused piece (a new data source, an AI summarization layer, an alerting flow, a single agent) is two to four weeks. A full application or niche product is six to twelve. We won't quote anything longer than twelve weeks without a checkpoint in the middle — long projects without checkpoints are how teams get stuck.
Real ones — an agent here means something that reads your systems, calls actual tools, and changes state, which is exactly why the boring parts get the attention. Every run leaves a readable trace, tool calls have permission and spend boundaries, and anywhere being wrong is expensive we put a human approval step in the path rather than trusting the model. We also build the eval suite over recorded runs, so you can tell whether a prompt change made things better or just different.
A whole one. Schema, screens, auth, billing, background jobs, deploys, error tracking — the parts that make it a product rather than a demo. That's the third line of work, and it comes in two shapes: greenfield applications built from the idea up, and setting up or integrating the systems you already run. If you need a native iOS or Android app, we'll build the API and refer you for the app itself.
Thirty days of follow-up support comes with every project — fixes, small adjustments, a follow-up review. Past that, we offer an hourly retainer at the standard rate. Monitoring systems need tending as sources change, and agents need it as models and tools move under them, so a light ongoing arrangement is common — but never an open-ended SLA.
Public registers and open data (USPTO, SEC EDGAR, SAM.gov, USAspending, and the like), third-party and partner APIs, bulk feeds, and your own databases. If a source is reachable and its terms allow monitoring, we can usually ingest, dedupe, and entity-match it. If it can't be done cleanly, we'll tell you before you commit.
Provider-agnostic — the latest OpenAI, Claude, and Gemini models, chosen per task and swappable behind a thin adapter. Every summary is grounded in the source record and linked back to it, and we build an eval harness so you can measure quality rather than trust a vibe.
Small changes (under a day of work) we just absorb. Anything bigger gets written up as a one-paragraph change order with a revised quote, which you approve before we start. No surprise invoices.
Only us. Tiny Prism is a one-person LLC — your data, your accounts, your records stay between us. We sign whatever NDA and data-processing terms you need; we also have defaults if that's faster.
Sometimes. If the badge in the top right says "taking briefs", we have room inside the next thirty days. Either way, we'll quote a realistic start date when you send the brief.
We read every brief and respond within one business day with either a yes, a no, or a referral to someone better suited.