I build AIthat holds upin production.
Muhammad Bilal is an applied AI consultant based in Germany who builds agentic AI systems that deliver measurable outcomes for clients — from strategy and architecture through to production delivery. The hub for client engagements, agentic tooling, and research residue. The personal half lives at bilalm.me.
Most AI work
never reaches production.
of enterprises report zero measurable profit impact from AI adoption. The shortfall is not in the models. It is in the engineering, the operations, and the discipline required to make a model survive contact with a real organisation.
I work in the gap between a working notebook and a system someone else can run on a Monday morning without calling me. That gap is where most projects quietly die — and where the only economics worth quoting actually live.
Three that shipped.
Selected from sixteen — chosen because the receipts are countable, recent, and attributable. The remainder live at /projects.
- i.
A BI stack that replaced a controlling role.
Ten weeks, settled as an unpaid PoC, turned into a paid engagement when the numbers held. A Power BI stack on top of Toggl, Absence.io, and a Node middleware bridging the auth protocols product teams pretend aren't there. Eight dashboards, twenty-three KPIs live by the close of phase I.
Read the engagement noteBILLABLE EFFICIENCY49%→65%Δ +16pp · Feb → Mar 2026TRACKED HOURS1,247this month · 23 projectsUTILIZATION78%target 80%DASHBOARDS LIVE8 / 823 KPIs · phase I closeDAILY BILLABLE EFFICIENCY · 22 WORKING DAYSFEB ←→ MAR 202614:02 · KPI feed refreshed · 23 of 23 ok14:02 · Toggl sync · 1,247 hours14:02 · Absence.io sync · 12 recordsFIG. p. 01 · Operations dashboard · project team · redacted facsimileAuthored schematic - ii.
The extraction behind STOXX-600 climate research.
Six reporting years of European annual reports rendered into a machine-readable corpus, with provenance back to the source page preserved. The downstream paper, on climate disclosure, runs the dataset through ClimateBERT. The acknowledgement is a small line of type. It has opened doors.
Read SSRN 4763140$ python scrape.py reports/STOXX600/2023/[OK] Allianz_AR_2023.pdf 247p 12,431 ¶[OK] BASF_AR_2023.pdf 312p 18,802 ¶[OK] Siemens_AR_2023.pdf 428p 24,113 ¶[OK] Nestle_AR_2023.pdf 356p 19,447 ¶[OK] Unilever_AR_2023.pdf 298p 16,221 ¶[OK] AstraZeneca_AR_2023.pdf 401p 21,884 ¶[..] HSBC_AR_2023.pdf …corpus.jsonlUTF-8 · NDJSON{"firm":"ALV","year":2023,"page":1421,"text":"…climate-related risks…"}{"firm":"BAS","year":2023,"page":2042,"text":"…Scope 1 emissions of 2.1 Mt…"}{"firm":"SIE","year":2023,"page":3318,"text":"…transition plan aligned with…"}{"firm":"NES","year":2023,"page":2877,"text":"…physical risk exposure…"}{"firm":"UNA","year":2023,"page":1944,"text":"…carbon price assumption €85/t…"}{"firm":"AZN","year":2023,"page":3722,"text":"…CapEx earmarked for…"}COVERAGE BY REPORTING YEAR3,484 OF 3,600 FIRM-YEARS · 96.8%201896%201994%202091%202195%202293%202386%PROVENANCE PRESERVED · PAGE-LEVEL · CITED IN SSRN 4763140DOWNSTREAM → CLIMATEBERTFIG. p. 02 · Corpus extraction pipeline · annual reports → JSONLAuthored schematic - iii.
Meeting notes to a fileable Jira backlog in five minutes.
Planning meetings produce transcripts. Jira expects epics, stories, acceptance criteria, and estimates. Between the two lies a tedious hour. This tool collapses it to about five minutes — parse, draft, file against the Atlassian REST API in a way any team lead can audit.
See the repositoryFIG. p. 03 · Transcript → parse → backlog · MIT, auditableLive capture · 25s loop
How I engage.
- i.
Small teams.
I work directly with the person who can decide. No account managers. No project triangles. The shortest line between a problem and a working system.
- ii.
Fixed scopes.
Six to fourteen weeks, written down before we begin. No retainers. No scope drift. No surprise invoices. The deliverable is on the page from day one.
- iii.
Production targets.
Every engagement closes on a measurable outcome — a number on a dashboard, a paper that cites the work, a system someone else now runs without me.
- iv.
Honest no.
If the project is wrong for me, or the model is wrong for the project, you'll hear it in the first call. The reply is part of the work.
Bring me a problem.
Send the constraint, the deadline, the shape of the data. I’ll tell you in the first call whether I’m the right person for it. The reply, as ever, is part of the work.
mbilal.workmail@gmail.com- /thisisbilal
- /bilal-m
- bilalm.me
- CET — UTC +1 / +2
- within 48 hours