fake_progress_

fake
progress

>ai agents (that finish) >ml on your data >full stack >architecture rescue >startups >vibe coding >qa & load >scrum (no theatre)
// our team
0+
core engineers · we scale to 10 on demand
// proven delivery
0+
production projects shipped — and counting
// always reachable
0h
24-hour team access while your project is live
// time-tested
0+
years on the market · same core team

About the name.

// why we called the company that

For years we've been joking about it inside the team. Every time a project needed a loader or a progress bar that doesn't really do anything, somebody would yell across the room — "just ship the fake progress". Hallucinated citations. Confident wrong answers. Bars that fill while nothing's actually happening. We've been laughing at it since long before there was a paper.

Then Carnegie Mellon turned our running joke into a benchmark.

REF: TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks — Xu, Neubig et al., Carnegie Mellon University, 2024–2025. arxiv.org/abs/2412.14161
setup
a simulated software company on open-source tools (GitLab, OwnCloud, Plane, RocketChat). 175 real tasks across software engineering, project management, HR, finance and admin. Agents browse, write code, run programs, and talk to 16 simulated coworkers.
scope
twelve frontier models tested — the best from Anthropic, Google, OpenAI, Meta, Amazon, Alibaba. Real money, real API costs, run with the OpenHands agent framework.
// full task completion · 175 tasks
Gemini 2.5 Pro · Google
30.3%
Claude 3.7 Sonnet · Anthropic
26.3%
Claude 3.5 Sonnet · Anthropic
24.0%
GPT-4o · OpenAI
8.6%
Llama 3.1 405B · Meta
7.4%
Amazon Nova Pro · Amazon
1.7%
The best agent on the planet finished 30% of ordinary office work. The rest — roughly 70% — ended in partial or outright failure. On well-scoped tasks. The researchers named the dominant failure mode "self-deception": agents inventing shortcuts that look complete but quietly skip the hard part.

fake progress

So we named the company after the joke. It's a contempt-flag. Every time you see our name on an invoice or a Slack thread, you're seeing what we refuse to ship — the bar that fills while nothing happens. we ship the opposite.

What we do

// services · 8 things we ship

Each tile below is something we've actually shipped this year. We don't list things we haven't built.

[ 01 ]

ai agents & chatbots

Agents and chatbots tailored to your operations — support, document drafting, internal assistants, escalation flows. Built with proper evals so they fail loudly, not silently.

[ 02 ]

ml on your data

Custom models trained on your own corpus — classifiers, scoring engines, recommendation, anomaly detection. Monitored, versioned, retrainable. No black boxes.

[ 03 ]

full-stack web

Web platforms end-to-end — UI, API, data, infra. Python, React, Postgres, AWS. Every system ships with tests, monitoring, and a runbook you can hand off.

[ 04 ]

architecture rescue

Inherit a system that's choking and bring it back to life — refactors, redesigns, migrations. We take over projects other teams abandoned and ship them to production.

[ 05 ]

startups: mvp → prod

From idea to first paying customers and beyond. Senior engineers from day one — without the cost of a full team. Equity-friendly engagements for the right founders.

[ 06 ]

vibe coding

AI-paired delivery for greenfield products and rapid prototypes. Several times faster than a traditional team — but with tests, docs, and observability. Same quality, half the timeline.

[ 07 ]

qa & load testing

Real QA — manual, automated, and load. Find what breaks before customers do. Regression suites, load profiles tuned to your real traffic shape, and honest verdicts.

[ 08 ]

scrum & delivery

Honest Scrum and Kanban — real sprints, real risk logs, real burndowns. We run the delivery; you keep strategic control. No standup theatre.

Last projects

// 5 of 10+ shipped · latest first · 2025–2026

A selection from our recent work. Real URLs where the work is public, real numbers from production.

PROJECT 01 :: visitweb live · 9+ years

VisitWeb — ad network with AI anti-fraud

100M+ ad shows/day·8 formats·9+ years live·our co-lead built it, sold it, still ships into it
problem

Ad networks bleed budget to bot clicks. ~12% of paid traffic was fake, refunds ran weeks behind the spend, and advertiser NPS sat at 6.

solution

A real-time AI verdict engine between every click and the invoice — 14 signals, 20M events/day, decided in 50ms, fakes blocked before they count.

results
~12% → ~2%
bot traffic share
-70%
advertiser refunds, y/y
6 → 9
customer nps, one quarter
100M+ ·
ad shows per day
PROJECT 02 :: redbus pilot · scaling

RedBus — route optimization for delivery fleets

logistics · fleets · last-mile·scaling 10 → 700 vehicles·commercial launch Q2 2026
problem

Planners burned 3–4 hours every morning routing by hand. Every order change forced a full re-plan, and fuel costs kept drifting up with no visibility.

solution

An engine that ingests orders, constraints and live traffic to produce routes in minutes — re-planning automatically, same code from 10 to 700 vehicles.

results
15 min
daily planning, was 3–4 hrs
~$62k ·
per fleet/year saved
10 → 700
vehicles, same codebase
Q2 2026 ·
commercial launch
PROJECT 03 :: caemap acceptance testing

CAEMap — Roerich Expedition Atlas

2,000+ artifacts · 5 of us·caemap.com · caemap.online
problem

Roerich’s 1924–28 expedition left 2,000+ artifacts scattered across archives with no public access point — and the previous contractor walked off mid-build.

solution

We rebuilt the platform from scratch: an interactive map pinning every artifact to where it was made, with 3D landscape views — shipped for the 100-year anniversary.

results
2,000+ ·
artifacts, geo-anchored
100 yr ·
anniversary, on time
5 of us ·
full delivery
2 sites ·
com · online
PROJECT 04 :: scorebazooka live

ScoreBazooka — is this YouTube video worth your time?

AI Chrome extension · 2 of us·scorebazooka.com
problem

YouTube is full of clickbait, AI-generated noise and confidently-wrong content — with no fast way to judge a video before you spend the time.

solution

A Chrome extension that grades any video on six independent axes via specialized classifiers — truth, relevance, ad-density, ethics, uniqueness, human-vs-AI — in one click.

results
6 axes ·
truth · ethics · human · more
1 click ·
full report, no setup
Live ·
public extension
2 of us ·
end-to-end build
PROJECT 05 :: kunama live

Kunama Translator — a low-resource language nobody else supports

low-resource language · iOS + Android · 2 of us·kunamaapp.com
problem

Kunama — spoken in Eritrea and Ethiopia — has zero support in any major translator and no public training corpus. Speakers lose access to tools everyone else takes for granted.

solution

We built a training corpus with native speakers, trained a model from scratch, and shipped a native offline iOS + Android app with flashcard learning built in.

results
1 language ·
unsupported by major ai
Offline ·
works without connectivity
iOS · Android ·
native, both stores
Live ·
kunamaapp.com

More we've shipped

Chess Academy Online — interactive chess school for beginners. AI sales-call quality control — auto-checks 100k+ calls/month for compliance. RAG agents on enterprise data — search and chat across messy Excels, ERP exports, and knowledge bases. Custom Telegram bots — onboarding, sales, ops, AI assistants for B2B teams. And more under NDA — ask us.