Based in Berlin · Available for AI engineering work

ArchitectingintelligentSystems.

I build AI systems and the evidence behind them: a phone receptionist in pilot stage that answers in German and English, an agent fleet that repairs itself overnight, and retrieval pipelines with dated eval results. CS student in Berlin.

View work Voice demo on request

01 — Proof

The difference between a demo and a system is measurability.

38%

Token cost per message, cut and measured

› 1,118 → 695 tokens, measured 2026-04-27 · hybrid RAG replaced prompt-stuffing

2.4%

Hard-failure rate, 1,000 agent runs from July to September 2026

› An autonomous agent fleet, 2,000+ runs since June, as of 2026-10-01 · lease-queue dispatcher, self-healer · a 43% failed-run backlog drained to 0 on 2026-08-22

12 of 12

Voice receptionist outcome eval, golden calls

› offline rubric, 2026-09-02 · pilot stage · per-call compliance evidence logged since 2026-08-23

Voice demo

Don't take my word for it — hear it.

The receptionist answers in German or English, books a table or an appointment, and handles interruptions without losing the thread. It tells you it's an AI in the first sentence (EU AI Act Art. 50) and records no audio, only a text transcript. The disclosure is logged per call in a tamper-evident journal.

The voice agent is in pilot stage, so its line is not public. Email me and I'll set up a demo call.

Request a demo call

Demo on request · you're talking to an AI

02 — Selected Work

AI Phone Receptionist

Pilot stage · line not public · German & English · Vapi

Owner-run businesses lose bookings to phones nobody answers. The demo assistant answers for a Berlin estate agency: it qualifies buyers and renters in English or German, saves the caller's details to the database at hang-up, and books through a tool webhook. It discloses that it's an AI in its first sentence, records no audio, and writes hash-chained compliance evidence per call — the parts a German business actually needs before it can use one.

Demo on request · 12 of 12 outcome eval (2026-09-02, offline) · per-call compliance evidence since 2026-08-23Vapi · Deepgram · ElevenLabs · Claude · FastAPI · PythonRequest a demo

Self-Healing Agent Fleet

Self-built · Operations

An autonomous agent fleet, each agent built for one job — code review, security sweeps, backup verification, evals, log digests. Active agents lease their work from a Postgres queue. A self-healer sorts every failed run into a failure class and requeues, skips or escalates it, no-LLM probes check the stack every five minutes, a cost-truth ledger enforces a daily cap, and a morning Telegram digest reports what happened. Before the healer shipped, 43% of runs were failed or stale; that backlog drained to zero on 2026-08-22.

Autonomous agent fleet · 2,000+ runs since June, as of 2026-10-01 · 2.4% hard failures in 1,000 runs, July to September 2026Python · PostgreSQL · PM2 · Claude Code

Multilingual RAG Commerce Agent

Pilot, paused · Commerce · WhatsApp

A WhatsApp sales assistant over a furniture catalogue, built for a business as a pilot, now paused. Hybrid retrieval — a keyword pass catches SKUs that embeddings dilute — replaced dumping the catalogue into every prompt. A semantic cache (95% cosine, 7-day TTL) absorbed repeat questions, and responses were cost-traced in Langfuse until it was parked in September 2026. It answered in German and English.

38% token cost cut (1,118 → 695 tokens per message), measured 2026-04-27Python · FastAPI · ChromaDB · Claude · Meta WhatsApp API

Autonomous Job Pipeline

Self-built · Automation · runs daily via the fleet

Two passes. The first runs daily via the fleet: an agent scores fresh Berlin postings, checks the matches and sends a Telegram digest, degrading gracefully when a source is down. On demand, a deeper pass pulls 83 company career boards directly — Greenhouse, Ashby, Lever, Personio, SmartRecruiters, Join — drops the roles I can't honestly apply to before it scores anything, checks the employer's own page is still open, then writes the application: a one-page CV tailored to the posting, and a letter an independent reviewer pass has to clear at zero unsupported claims. It found the role you might be hiring for.

2,981 postings from 83 career boards in the 2026-09-26 sourcing run · no letter shipped with a claim my CV can't backPython · Claude · Firecrawl · cronGitHub

Evidence

Eval gates, run counts, and the sales pipeline.

12 of 12

Voice receptionist — phase-7 outcome QA (golden calls)

2026-09-02 · pilot stage

› Offline, deterministic rubric

10 of 10

Restaurant bot — RAG retrieval (gold questions)

2026-08-25 · offline

› 10 gold-standard retrievals, hybrid search, no LLM calls

Sales OS

Qualifies local businesses for the voice receptionist

Lead pipeline · manual runs

› Finds businesses on Google Places, scores how many calls each one is likely to miss, and keeps outreach consent-first: no cold messages (UWG §7)

2,000+

Agent runs since June

2026-10-01 · Autonomous agent fleet · Postgres lease queue

› A self-healer classifies every failed run and requeues, skips or escalates it

03 — How I Build

How I build systems that ship.

Understand
Define the real problem and what “working” means in numbers before writing code.
Architect
Design retrieval, agents, data flow, and failure modes together — not failure modes last.
Build & evaluate
Ship with an eval gate in CI, so quality is proven on every commit, not claimed once.
Observe & iterate
Measure real usage and real cost after launch; tune on what the numbers show.

04 — Stack

The tools I reach for.

AI & retrieval

ClaudeChromaDBhybrid searchsemantic cachingeval gates

Backend & data

PythonFastAPIPostgreSQLSupabaseRedis

Voice & messaging

VapiDeepgramElevenLabsWhatsApp Cloud APITelegram

Infra & observability

DockerPM2CloudflareGitHub Actions

Frontend

Next.jsTypeScriptThree.jsTailwind

Quality

pytestruffGitHub Actionsgitleaks

05 — About

The person behind the systems.

I'm Shehryar, a computer science student at Arden University in Berlin. For the past six months I've built and operated an AI stack on my own server: Linux and Docker at the base, RAG and eval layers in the middle, and the frontend you're reading now on top.

I work eval-first. What separates a demo from a system is evaluation, tracing, and cost, so every project here ships with a dated eval result behind it. When something breaks, I fix the cause rather than the symptom. A bot token once showed up in a log; I fixed the logger, not just the key.

When a product needs to speak both English and German, I build that too.

Based in Berlin · Available for AI engineering work.

Currently

Studying: BSc Computer Science, Arden University
Focus: AI systems · Full-stack
Availability: Based in Berlin · Available for AI engineering work

06 — Contact

Let's build something that works.

Tell me what you're building, or just say hello.

Or directly

ArchitectingintelligentSystemsSystemsSystemsAgents.