Back home
WORK
Proof, not promises.
Systems we've built inside real operations — not demos, not pilots that never shipped. Here's what the problem was, what we built, and what changed.
- 01
Social media
Agentic personas that show us what content is actually working
- Problem
- How do you know what kind of content is actually landing with a real audience — not what you think should work, but what people are genuinely scrolling to and watching?
- Approach
- We built agentic personas — AI agents that behave like real audience segments, not just demographic labels — and set them loose to scroll feeds and watch reels the way a real person would. Instead of asking a persona to react to something we hand it, we let it browse on its own, then analyze what it actually stopped on, watched, and engaged with. That surfaces the real patterns — the topics, formats, and styles a given audience segment is genuinely drawn to — instead of guessing from a demographic profile.
- Outcome
- A strategy team gets a real read on what content resonates with a given audience segment, grounded in actual browsing behavior — not assumptions about what that audience "should" like.
- 02
Logistics
Real-time fleet intelligence + predictive load matching
- Problem
- A national trucking network had no single view of itself. Dispatchers didn't know where trucks actually were, which ones were genuinely idle, or which routes were quietly losing money — matching a load to a truck was mostly guesswork and phone calls.
- Approach
- We built a live intelligence layer on the network's own vehicle-movement feeds. A location ping alone tells you almost nothing, so we modelled each vehicle's historical movement sequence to infer its actual behaviour — is this truck running a fixed shuttle between two hubs, or a genuine one-off long haul? That turned a raw location feed into a real map of recurring routes and corridors. On top of that we built a load-matching engine that weighs a truck's inferred route pattern alongside its capacity and five other factors, plus a profitability calculator that works out margin per route, not just per trip.
- Outcome
- Dispatchers now see which trucks are genuinely idle versus mid-route on a habitual shuttle — and loads get matched against a truck's real behavior instead of its last known GPS ping. A truck running the same corridor every week finally gets matched like one, instead of being re-evaluated from scratch on every load. Utilization and margin are visible per route, not just per truck.
- 03
Manufacturing
Agentic quoting platform built with real machine knowledge
- Problem
- Quoting a complex piece of industrial machinery isn't one number — it's dozens of configurable rows: parts, specs, options, each with its own price and dependency on the others. Building that by hand from a customer's plain-language brief took a sales engineer days, and any mistake in translating "what the customer asked for" into "the right configuration" meant a wrong quote going out.
- Approach
- We built an agentic platform that actually knows the machine — its full configuration logic, every valid part combination, every pricing rule — not a general assistant bolted onto a form. Fed a customer's brief, it reasons through the request and builds the entire multi-row quote against that real product knowledge, row by row, the way an experienced sales engineer would. A sales engineer can still question or adjust any single row and the platform reasons over just that line, in context, without re-deriving the whole quote. The one thing it never touches is arithmetic: every price is pulled from a fixed, deterministic product configuration, so the agentic reasoning handles interpretation and the math stays fully traceable. Once approved, the quote flows straight into the client's SAP system as a sales order — no re-entry.
- Outcome
- A quote that took days of manual, line-by-line work now takes minutes — built by a system that actually understands the machine, not just the form.
- 04
Energy
A chat interface that reasons over a real financial model — not a chatbot bolted onto Excel
- Problem
- A renewable energy financial model — the kind used to underwrite a solar or wind power purchase agreement — isn't one formula, it's 50 to 100+ interdependent assumption cells feeding into a handful of final numbers: NPV, IRR, LCOE. Understanding why a number moved, or testing "what if the PPA price dropped to $40/MWh," meant an analyst manually tracing formulas cell by cell through a maze only the person who built the model could really navigate.
- Approach
- We mapped the entire model as a dependency graph in Neo4j — every input, every intermediate calculation, every final metric, connected the way it actually feeds through the spreadsheet. That's what lets a question like "why is NPV this value" or "what inputs actually affect IRR" get answered by tracing real structure, instead of re-parsing formulas from scratch every time. But a graph alone can explain relationships — it can't give you a trustworthy number. So we wired it directly to a live Excel engine, which recalculates the actual model and returns Excel's real computed output, not a reimplementation of Excel's math that could quietly drift from the real thing. On top of both, we built a scenario engine that handles natural-language what-if and goal-seek queries directly — "find the minimum PPA price for a 12% IRR" — without anyone touching a cell.
- Outcome
- A chat interface replaces clicking through dozens of interdependent cells — ask the question, get an answer that's traceable to the real dependency structure and verified against Excel's own math, not an approximation.
CONTACT
Tell us what's actually broken.
Thirty minutes, one of the founders, no sales team in between.