Engineering Practical AI & Developer Tools: Voice, Image Pipelines & QR

Why we build infrastructure instead of AI wrappers

Most software teams talking about AI and developer tools today are gluing a thin chat interface onto a single OpenAI endpoint and hoping nobody looks too closely at the latency, the unit economics, or what happens when real production traffic arrives.

When you run engineering for high-volume platforms in Southeast Asia and Europe, thin wrappers break immediately. Phone calls drop when voice synthesis lags by 800 milliseconds. Image CDN bills triple overnight when a property portal generates responsive WebP variants for two million photos. Printed QR codes on packaging fail to scan because a designer pasted a PNG logo over the center timing modules.

Inside Brainstorm IT, whenever we hit an infrastructure problem that off-the-shelf tools either overprice or mishandle, we engineer a clean, deterministic solution and turn it into a reusable product. Here is a technical teardown of three systems we built in-house: SAGOT (Taglish AI voice telephony), Viucraft (high-throughput image processing), and QRlfy (GF(2) mathematical branded QR codes).

---

1. SAGOT: Engineering a low-latency Taglish AI voice agent for Philippine telephony

If you have ever tried to deploy an American AI voice agent to handle customer calls in Manila, you know why it fails within the first thirty seconds.

Filipino professionals and consumers rarely speak pure textbook English or pure formal Tagalog on the phone. They speak Taglish, switching seamlessly between English and Filipino inside the same sentence ("Yes, available pa yung schedule tomorrow afternoon, pero can you send the details sa Viber?"). When a standard US voice pipeline hears that sentence, three things go wrong:

1. Transcription hallucination: English-only speech-to-text models mangle the Tagalog words into phonetic English nonsense, feeding garbage into the LLM.
2. Awkward turn-taking latency: If the voice pipeline waits for a massive frontier LLM to reason before starting speech synthesis, the caller hears two seconds of dead air, assumes the line dropped, and says "Hello? Nandyan ka pa?", triggering a barge-in collision.
3. PSTN telephony friction: Cloud voice demos work fine in a browser WebRTC tab, and fall apart when routed through real Philippine mobile carriers and landline trunks.

How we architected SAGOT

We built SAGOT (sagot means "answer" in Filipino) from the telephony layer up to handle real PSTN calls in Philippine English and Taglish:

• Direct SIP termination on Manila DIDs: Calls route through dedicated Philippine local numbers (+6328 Manila landlines) terminated into our own FreeSWITCH media server running in Singapore (ap-southeast-1) to keep network hops across Southeast Asia minimal.
• Bilingual STT + sub-second LLM orchestration: Audio streams into Whisper speech-to-text tuned for bilingual Filipino-English transcripts, which feeds a tightly scoped Claude Haiku conversation state machine. Using a fast, structured model keeps time-to-first-token low enough for natural phone cadence.
• Low-latency multilingual voice synthesis: Responses stream through ElevenLabs (eleven_flash_v2_5) with localized pronunciation handling so currency amounts in Pesos, Manila barangay and street names, and conversational Taglish acknowledgments sound natural rather than robotic.
• Structured post-call extraction: Every completed call automatically generates a structured summary, sentiment tag, and action ticket pushed straight into the operator's CRM or operations board.

For Philippine businesses in real estate brokerage, clinics, logistics confirmation, and BPO qualification, this turns outbound follow-ups and inbound triage from a staffing bottleneck into a reliable API.

---

2. Viucraft: A predictable alternative to credit-based image CDNs

When we built and scaled property portals, where a single marketplace handles millions of high-resolution listing photos uploaded from agents' phones over patchy mobile connections, we learned the hard way how quickly image processing costs can spiral. (We cover the full architectural blueprint in handling millions of property photos: storage, CDN, and image pipelines.)

Developer teams usually start with Cloudinary or imgix because URL-based image resizing is convenient. Then traffic grows, modern SEO demands serving multiple responsive sizes in both WebP and AVIF, and the monthly bill suddenly jumps from a few hundred dollars to several thousand because every variant and bandwidth gigabyte burns "credits."

How we architected Viucraft (viucraft.com)

We built Viucraft to give engineering teams the exact developer ergonomics they love about URL-based image CDNs, without punitive credit math:

• libvips at the core: Instead of memory-heavy ImageMagick processes that spike RAM on large uploads, Viucraft runs on a custom libvips processing core (php8.3-apache-vips), streaming pixel transformations with minimal memory footprint and multi-core speed.
• Per-user isolated subdomains: Each customer account gets its own dedicated delivery subdomain, cleanly isolating cache invalidation, origin rules, and access keys.
• Drop-in JavaScript SDK (viucraft-js) and CLI: Frontend teams drop in our open-source JavaScript SDK to generate responsive srcset URLs, crop modes, quality optimizations, and format negotiation on the fly, migrating off legacy image CDNs in an afternoon.

---

3. QRlfy: Branded QR codes via GF(2) linear algebra, not broken image overlays

Every consumer brand, event platform (including our own wedding platform, Ceremo), and restaurant chain wants branded QR codes that actually look like their brand rather than an ugly black-and-white industrial barcode.

Today, companies choose between two flawed hacks:

1. Center Logo Overlays: Pasting a square logo over the middle of the QR code. Technically, this deliberately damages up to 30 percent of the QR code's data modules and relies on Reed-Solomon error correction (Level H) to survive the damage. Because the error-correction budget is already spent by the logo, the moment the printed code has a slight smudge, crease, or dim restaurant lighting, phones fail to scan it.
2. Diffusion AI QR Art: Running a QR code through Stable Diffusion ControlNet to make it look like a painting. These look impressive on social media, require multiple seconds of careful camera aiming to decode, and are non-deterministic (change one character in the URL and the visual art changes completely).

How we architected QRlfy's artcode engine

We built QRlfy around a pure mathematical approach inspired by Russ Cox's QArt algorithm, implemented from scratch in a deterministic TypeScript engine (ISO 18004 byte mode + Reed-Solomon encoder) paired with a Python validation harness:

• Sculpting padding bits via Gaussian elimination over GF(2): In a higher-version QR code carrying a clean short URL, a large portion of the matrix consists of free padding bits coupled to Reed-Solomon parity bits. Because Reed-Solomon over GF(2) is linear, we run Gaussian elimination over GF(2) to choose the exact free bits that force the visible QR modules themselves to match the contrast map of your brand logo.
• Zero error-correction budget spent: Every single dot forming the logo is a valid, mathematical data module in the QR specification. The code retains 100 percent of its Reed-Solomon error-correction capacity for real-world print wear and camera glare, while the encoded payload remains a clean URL with zero #junk fragment hacks.
• Deterministic and benchmarked across 196 brands: The same logo plus the same URL always produces a byte-identical PNG with zero machine-learning randomness. In our automated decode benchmark against a 196-brand test suite using zxing-cpp and OpenCV across five progressive distortion rungs, 196 out of 196 codes passed every single decode rung with zero per-logo manual flags.

Why this matters when you choose an engineering partner

Whether you are a customer looking to automate operations with AI, a CTO looking for developer tooling that respects your unit economics, or an investor evaluating our bootstrapped venture studio in Southeast Asia, these three products illustrate how Brainstorm IT works.

We do not stop at surface-level integrations. When a problem calls for configuring FreeSWITCH SIP trunks in Singapore, tuning libvips memory pipelines, or writing a GF(2) linear algebra solver in TypeScript, we have the senior engineering depth to build it right and verify it with automated benchmarks.

If you have a hard technical problem in voice AI, media pipelines, or custom operational software, start with our two-week Technical Deep Dive or book a 30-minute call with Dan.