Shujaan Azhar

Four things I actually do, each one demonstrated below rather than described.

1. 01 Development Open source 2026 Solo

   TaxBacha

   A tax optimiser for Pakistani salaried filers — not “what do I owe”, but “what could I owe instead”.

   Befiler and FBR IRIS optimise for filing compliance. TaxBacha optimises the bill: it computes your liability against the FY 2025-26 salaried slabs, then ranks the legal instruments that reduce it — VPS under §63, Zakat under §60, approved donations under §61 — by rupees saved. No API keys, no account, no salary leaving the machine.

   Slabs

   FY 2025-26

   Keys required

   none

   Data leaves device

   never

   • Next.js 16
   • React 19
   • TypeScript
   • Tailwind 4
   • shadcn/ui

   Read the source ↗

   tax-engine.ts FY 2025-26 · salaried

   Annual salary

   PKR 4,800,000

   0%

   1%

   11%

   23%

   30%

   35%

   six brackets · 9% surcharge above PKR 10M

   Tax as filed 861,000

   →

   Optimised —

   Tax saved —

   Zakat§60 VPS§63 Donations§61

   bracket reached unreached Zakat deducts from income; the rest are credits at your average rate.

   This is the real engine. The slab function and every credit rule below are the exact logic from app/lib/tax-engine.ts, ported line-for-line. Drag the salary.

2. 02 Voice & Realtime Open source 2026 Solo

   VICIdial AI Voice Agent

   A call-centre voice agent with barge-in that runs entirely on my own GPU. No cloud, no carrier, no PSTN.

   Speech recognition, the language model, and speech synthesis all run locally; the “customers” are SIP softphones on a private subnet. It hooks into Asterisk audio over AudioSocket — an 8 kHz PCM TCP stream — and VICIdial’s campaign, lead and reporting layer wires in on top. Total running cost is zero and no audio ever leaves the machine.

   Response latency

   ~0.3 s

   Running cost

   $0

   Built on

   RTX 4060 · 8 GB

   • faster-whisper
   • Ollama · qwen2.5:7b
   • Piper TTS
   • Asterisk 18
   • KVM/libvirt
   • Python 3.12

   Read the source ↗

   deployment topology zero PSTN

   Host 192.168.122.1

   • VAD energy · adaptive floor CPU
   • STT faster-whisper small.en GPU
   • LLM Ollama qwen2.5:7b GPU
   • TTS Piper lessac-medium CPU

   AudioSocket

   8 kHz · 16-bit · mono PCM

   KVM guest 192.168.122.10

   • Asterisk 18 SIP / RTP · AudioSocket
   • VICIdial campaigns · dialer daemons
   • MySQL + Perl lead database
   • Apache / PHP agent web UI

   libvirt NAT bridge virbr0 · 192.168.122.0/24 · no carrier, no DID, no cloud API

   GPU CPU round trip ~0.3 s with barge-in cost $0

   The actual deployment topology — host and guest on one libvirt NAT bridge, no PSTN anywhere in it. The pulse is a turn of the audio loop.

3. 03 Voice & Realtime Open source 2026 Solo

   Jarvis — clap-activated assistant

   Double-clap; it wakes, listens, answers out loud, and goes back to sleep. Every hosted hop has a local fallback.

   Clap detection runs continuously on MFCC embeddings matched against a voice-print you calibrate yourself, so it fires on your clap and not a door closing — without holding the mic open for transcription. It keeps a persistent memory of you in a plain-Markdown Obsidian vault, reads your calendar, and hands code changes to Claude Code, routed to the right repo by spoken alias.

   Wake word

   your own clap

   Local fallbacks

   STT + TTS

   Memory

   Markdown vault

   • Python 3.12
   • MFCC / librosa
   • Groq
   • faster-whisper
   • Piper
   • Claude Code

   Read the source ↗

   degradation path network up

   hosted local

   1. Wake — MFCC voice-print always on device
   2. Hear Groq Whisper faster-whisper falls back
   3. Think Groq LLM · Claude — needs network
   4. Speak Groq Orpheus Piper falls back
   5. Remember — Markdown vault always on device

   Clap detection runs continuously against a voice-print you calibrate yourself — it fires on your clap, not a door.

   active path standby press network up to pull the plug

   Hosted path on top, local fallback underneath. Toggle the network to watch it degrade instead of die.

4. 04 Voice & Realtime Open source 2026 Solo

   Decepticon — autonomous meeting agent

   Joins your Google Meet calls from your calendar, listens, decides whether to speak, and replies in your cloned voice.

   The hard part is not speaking — it is knowing when not to. Transcription runs locally on GPU with Faster-Whisper, a local Mistral decides whether the moment warrants a reply, and Chatterbox synthesises the answer with zero-shot cloning from a ten-second reference clip. Audio moves over virtual PulseAudio devices, so there is no physical mic or speaker in the loop.

   Runs

   fully local, on-GPU

   Voice clone

   zero-shot, ~10 s ref

   Hardware needed

   none

   • Faster-Whisper
   • Ollama · Mistral
   • Chatterbox TTS
   • PulseAudio
   • CUDA

   Read the source ↗

   turn decision sample meeting

   1. Sara so who is picking up the migration this sprint hold
   2. Omar I can take the schema half of it hold
   3. Sara right — and the rollback plan? hold
   4. Omar Shujaan wrote that one last quarter hold
   5. Sara Shujaan, can you walk us through it? speak
   6. agent speaking — cloned voice tts

   turns heard 0 turns spoken 0 restraint is the feature

   hold speak Faster-Whisper → Mistral → Chatterbox, all on-GPU

   The decision the agent actually makes on every turn boundary. Most of them are “stay silent”.

5. 05 Agentic Systems Open source 2026 Solo · hackathon

   Sahulat — service orchestrator

   Seven agents that replace the thirty minutes of phone calls it takes to book a plumber in Pakistan.

   Type a request in Urdu, Roman Urdu, or English. Intake structures it, Discovery filters the directory by haversine distance in pure Python, Ranking scores candidates, and then the Coordinator spawns three concurrent negotiations — each provider played by an in-character LLM persona — asking can you come, what time, what price. It books the best committed offer. Built for the AI Seekho 2026 Antigravity Hackathon.

   Agents

   7

   Negotiations

   3, concurrent

   End to end

   ~5–8 s

   • FastAPI
   • Llama 3.3 70B
   • Llama 3.1 8B
   • Groq
   • Expo / React Native
   • WhatsApp Cloud API

   Read the source ↗

   coordinator agent sample run · 3 × Llama 3.3 70B

   rank score

   distance0.4 rating0.3 avail0.2 price0.1

   • Ahmed R. electrician · 1.2 km · 4.6★
     openquotecounterclose
     PKR— ETA— —
   • Bilal K. electrician · 2.8 km · 4.8★
     openquotecounterclose
     PKR— ETA— —
   • Kashif M. electrician · 0.9 km · 4.2★
     openquotecounterclose
     PKR— ETA— —

   coordinator waiting on all three…

   in flight committed 7 agents · end to end ~5–8 s

   The Coordinator’s three parallel negotiations, and the scoring rule that ranked the providers going in.

6. 06 Agentic Systems Deployed 2026 Solo · personal

   Sheikhspeare — Slack agent

   An always-on Slack agent in iambic pentameter that knows who is asking, and refuses the ones who shouldn’t.

   The interesting problem in a personal assistant that lives in a shared workspace is not personality — it is containment. Identity is resolved per message before any tool is reachable, so the private toolset and the owner’s context are gated at the boundary rather than guarded by a prompt. It runs 24/7 as a Socket Mode worker with an owner-only conversation log.

   Uptime

   24/7 worker

   Gate

   per-message identity

   Voice

   Shakespearean

   • Claude Agent SDK
   • Python
   • Slack Socket Mode
   • Railway
   • Docker

   identity gate per message

   both send “what did I say I'd ship this week?”

   @shujaan owner

   resolve

   owner

   toolset

   3 mounted

   • ✓ owner context mounted
   • ✓ private memory mounted
   • ✓ conversation log mounted

   @anyone-else guest

   resolve

   not owner

   toolset

   0 mounted

   • · owner context withheld
   • · private memory withheld
   • · conversation log withheld

   The gate runs before the model is handed a toolset — refusal is a property of the boundary, not a line in a prompt asking it nicely.

   The same question, asked by two people. Identity resolves before the toolset is even visible.

7. 07 Core AI Live in production 2025— Design & build · core engineer

   AI Lesson Plan Generator

   Curriculum in, classroom-ready bilingual lesson plans out — with a reviewer that can send them back.

   Multi-grade, English and Urdu, with auto-generated illustrations. The part that matters is the gate: an AI reviewer scores every plan against a pedagogy rubric before it ships, and a failing plan is regenerated rather than delivered. The async pipeline is migrating to a LangGraph state machine so each stage is inspectable and retryable on its own.

   Languages

   English · Urdu

   Quality

   gated, not assumed

   Every call

   traced

   • Python
   • FastAPI
   • LangGraph
   • Redis
   • Langfuse
   • AWS EC2

   generation pipeline LangGraph state machine

   1. intake curriculum + grade
   2. draft plan, EN + UR
   3. figures illustrations
   4. review pedagogy rubric
   5. deliver to the classroom

   below threshold → regenerate

   rubric

   • objectives
   • Bloom's coverage
   • standards alignment
   • assessment for learning
   • bilingual parity

   idle

   in flight passed the gate a failing plan is regenerated, not shipped

   The generate → review → gate loop. A plan below threshold does not ship; it goes back around.

8. 08 Core AI Live in production 2025— Design & build · core engineer

   AI Exam Generator

   Any teaching content becomes a standards-aligned exam — as structured data other platforms consume over HTTP.

   Multiple question types, bilingual, with figures. Because it returns structured exam data rather than a rendered PDF, it works as a reusable content API instead of a one-off document maker. A fourteen-check, standard-mapped reviewer grades every exam on a calibrated scale before it is returned.

   Reviewer checks

   14

   Returns

   structured data

   Consumed by

   other platforms

   • Python
   • FastAPI
   • Redis
   • Langfuse
   • AWS EC2

   reviewer harness 14 standard-mapped checks

   exam under review

   • multiple question types
   • English + Urdu
   • figures
   • standards-mapped

   01 02 03 04 05 06 07 08 09 10 11 12 13 14

   —

   waiting for an exam…

   checking passed graded on a calibrated scale before the payload is returned

   The reviewer’s check surface. Every generated exam is graded against all fourteen before it leaves the service.

9. 09 Core AI Shipped 2026 ML engineer

   Vision Exam Parsing

   Swapped a GPT-4o vision call out of a production parser for a YOLO detector I fine-tuned myself.

   Handwritten exam pages had to be segmented and cropped at scale, and a hosted vision model was the wrong tool for a fixed, repetitive layout problem. I curated a domain dataset, fine-tuned YOLO v11/v12 over 100+ epochs, and paired it with OCR — self-hosted, so throughput is a function of my own hardware rather than someone’s rate limit. It ships behind FastAPI with APM tracing and process-pooled workers.

   Training

   100+ epochs

   Replaced

   a hosted vision call

   Now runs

   self-hosted

   • YOLOv12
   • Ultralytics
   • PyTorch
   • Google Vision OCR
   • FastAPI
   • Elastic APM

   YOLOv12 · fine-tuned self-hosted

   regions cropped → OCR

   • marks box 46 × 24
   • question 240 × 54
   • diagram 100 × 70
   • answer 240 × 70

   GPT-4o vision call → detector I trained

   100+ epochs on a curated domain dataset. Throughput is now a function of my hardware, not someone's rate limit.

   What the detector does to a page: regions found, classified, and cropped for OCR downstream.