ANNI is a family of AI systems that run on hardware you own, work in the local language, and don't depend on a paid cloud service. Designed and built by Kahel Estrada.
rSCENE 2026 AI Vibe Coding Challenge (Open Category). SAMMI went from problem statement to live site in 2 h 20 min at Tandaya Hall, Catbalogan.
GOLDEN THESIS
DLSU Golden Thesis Awardee for a fully functional IR-supported facial recognition system.
IEEE
Published researcher, IEEE (ICMIC 2025, Cebu). Military and maritime medical AI: a low-bandwidth, local medicine-assistance bot for deployed units, scoring an F1 of 0.73, above the 0.60–0.70 typical of medical dialogue systems. Read on IEEE Xplore → Military application
Named champion on stage at rSCENE 2026, Catbalogan City, October 2026.
SAMMI · reading Waray, Filipino and English price messages on site
Each member of the ANNI family is a specialist. ANNI manages them; each one does one job well.
1st place rSCENE 2026
Case file 001 · Catbalogan, Samar
SAMMI
Samar-ANNI · the Waray-Waray specialist
First deployment: Presyo Catbalogan, a citizen commodity-price platform. Sellers update prices by chatting with Sammi, a Visayan spotted deer, on Messenger in Waray, Filipino or English. A local AI model on an on-site edge AI device reads each free-text message into item, market and price.
Fully correct (32 test sentences)94%
Item and price correct100%
Time per message≈ 2.2 s
Cost per message₱0
Offline fallback (rule-based)91%
Checks every price against official commodity lists
Flags unusual prices for a market monitor to verify
3-day price forecast and buy-now-or-wait advice for residents
Built on open-source models, tuned for Waray and Filipino with validation, anomaly checks and an offline fallback
With the champion's medal, certificate of recognition and trophy, right after the awarding.A seller updates a price in Waray on Messenger. Sammi works out ₱300/kg from "3 kilo 900". Pilot demo.The live Presyo Catbalogan site opens with a guided tour by Sammi over an illustrated Catbalogan.
Built 7 Oct 2026 at Tandaya Hall, Catbalogan: problem statement to live site in 2 h 20 min. Individual entry. Pilot data: 39 commodities, 12 simulated stalls, 3 Catbalogan markets.
Where ANNI began: a personal assistant that runs entirely on a small edge AI device on the local network, managing a home lab.
Local AI model on an edge AI device; chat console on the local network
Long-term memory, so it remembers past conversations and decisions
Asks first: sensitive actions wait until a person taps Confirm
Speaks with on-device speech synthesis
Services restart themselves after a power or network loss, checked by 37 automated health checks
Can ask a cloud AI for a second opinion only with the owner's consent, with private details removed first
The ANNI console: chat on the left, the lab's live status on the right. Interface preview with sample data.On a phone. Before ANNI restarts anything, a "Needs your OK" card waits for a person to tap Confirm, with a countdown.
Golden Thesis
Case file 003 · DLSU
Real-Time Facial Recognition Headgear
Research · offline · works in the dark
Working prototype
A wearable that recognises faces in real time and learns new ones on the spot, running entirely on a microcontroller smaller than your thumb. No cloud, no internet. An infrared camera lets it work in a darkened room as well as in normal light.
99.3%accuracy
0false rejects
0.9994AUC (1 = perfect)
Test comparisons on the device137
Genuine matches accepted25 of 25
Impostor attempts refused111 of 112
LightingDark room (IR) + room light
Learns a new faceOn the spot
DLSU Golden Thesis Award. In collaboration with Yebes, M., Dominado, F.R. and Hortaleza, D.R.; Kahel built the system architecture and the on-device optimisation. Results from the on-device trial at the accuracy-optimal setting; the deployed setting trades a little accuracy (97.8%) to guarantee zero false rejections.
Paper in progress
Case file 004 · Research
Offline AI Tutors for Filipino Classrooms
Small language models · no internet needed
In July 2025, nearly 12,000 Philippine public schools still had no internet. This study asks whether a small AI tutor can run entirely inside the school, on one low-power edge AI device, and answer school-level science, maths and reading questions in English and Filipino.
Small language models tested4
School-level test questions2,600
Answers measured22,240
First word of an answer0.17–0.34 s
Power while answeringabout 10–12 W
Filipino reading accuracy, same 900 questions
General model, untuned63.1%
Regional model, untuned69.0%
Adapted on a laptop74.6%
−55%energy per correct answer
Hoursof training on a laptop
Samespeed, power, memory
A few hours of training on an ordinary laptop, using public Filipino data, lifted Filipino accuracy above every untuned model, with no change to the device. Energy per correct answer fell from 33.9 J to 15.1 J.
Honest trade-off, reported in the paper: the adapted models got worse at multi-step maths. Publication in progress.
In development
Case file 005
ANNI Quant
Market intelligence · ongoing project
In development
ANNI's market desk. Quant gathers forecasts from market prediction platforms through their data feeds and watches major news outlets, so the crowd's predictions and market-moving news land in one place for trading decisions.
Pulls live forecasts from prediction-market APIs
Monitors major news media for market-moving stories
Brings both together as signals for trading
Ongoing project: no performance numbers are published yet. Results will be added once they are measured.
Case file 006 · Education
AI Literacy Teaching
Online classes · generative AI for students and teachers
Active
Kahel teaches AI literacy online with Philippine Coding Camp: what generative AI is, how students can use it for learning and writing, and how teachers can use it to prepare lessons and assessments. The course speaks to a Waray audience and is built as a template: translate it and it fits any regional language or Filipino.
A lesson slide in Waray: "Bisan diin na an generative AI" (generative AI is everywhere now), linking students and teachers to the school.
Classified
Case file 007 · Restricted
Military Grade Intelligent System
Clearance required
Military applicationAdvanced researchNeural architectureRestricted access
Security log from a system test: forged, replayed, unknown and unencrypted messages are refused, valid ones accepted. Identifying fields redacted.From the training logs: 26 model training runs over two days, each tracked to its best checkpoint.
intelligent system for military applications, advanced research a multi-layer neural architecture autonomous
secure edge signal under restricted access.
Full file available to authorised parties on request.
Case file 008 · Automation
Custom Agent Skills
Works with Claude · Antigravity · ANNI
In daily use
Hand-written skills that turn a whole job into a repeatable playbook an AI agent follows step by step, with the safety checks built in. The same skill runs in Claude, in Antigravity and in ANNI's own system, so a workflow is written once and every agent does it the same way.
20custom skills
3skill sets
3agent platforms
---name: autobuild
description: Hands-off build. One idea in, a live verified app out: spec, database, repo, hosting, 3 features, polish, poster, demo script.
---
Rapid build kit (12 skills): idea to spec, scaffold, database and hosting, one feature at a time, checkpoints, polish, performance audit, deploy, poster and demo script. The kit behind SAMMI's 2 h 20 min build.
Video production pipeline (6 skills): studies a reference video's art style into a style guide, then runs script, original characters, settings and storyboard, low-poly 3D shots, and clean vector artwork.
Operations (2 skills): a council that syncs progress across several AI tools and hands off tasks, and a backup routine that only frees disk space after the upload is verified.
Hybrid by design: skills can run against a local AI model on site, with cloud models used only when approved.
One of Kahel's skills, shown above in its real format: plain text a person can read and check.
> FILE 02 · HOW ANNI WORKS
One manager. Many specialists. One rulebook.
ANNI is the manager: the central intelligence on your local network. Each family member is a specialist with one job. All of them follow ANNI's written rulebook, and a person approves anything sensitive.
Your local network · on site
ANNIManager · Autonomous Neural Network Interface
SAMMIWaray-Waray language Live
AssistantVoice & tasks Live
QuantMarkets & news Building
Rulebook + human approvalWhat each agent may do, what data it may touch, and what needs a person's OK
> ONE MESSAGE, START TO FINISH
How SAMMI handles a price update
01 · Seller"An alamuhan naman 3 kilo 900"sent on Messenger, in Waray
02 · Read on siteAlumahan · ₱300/kgthe local AI model turns free text into item, market and price, about 2.2 s
03 · CheckedAgainst the rulesofficial commodity list, the seller's own stall, and a flag if the price looks unusual
04 · Published"Salamat! Updated…"the price board and buy-now-or-wait advice update for residents
Runs on site
The AI model runs on an edge AI device at your location. Messages are read there, so there is no per-message cloud bill: SAMMI costs ₱0 per message.
Follows written rules
Every ANNI agent works from a written rulebook: its job, its limits and its language. Rules can be read and checked by people.
Asks before acting
Sensitive actions wait for a person to approve them. The ANNI Assistant already works this way: nothing runs until someone taps Confirm.
Keeps going
If the AI device is offline, SAMMI falls back to a rule-based reader (91% correct) instead of stopping. The Assistant's services restart themselves after a power or network loss.
Planned
> WHERE ANNI IS GOING
ANNI as manager of a team of agents: ANNI assigns work to specialist agents, checks it against the rulebook and reports back.
ANNI's own language model: a small model being designed and built from scratch, to handle routine work on site.
These are plans in progress, not finished products.
> FILE 03 · SERVICES
Bring ANNI to your office
Every project starts with a free consultation: tell me the problem, and I'll tell you honestly whether local AI is the right fix.
Service 01
Agentic systems
AI agents that take on real tasks under a written rulebook, with a person approving anything sensitive. The way ANNI manages her specialists.
Service 02
Local-language AI assistants
Chatbots and assistants that understand Waray, Filipino and English, on Messenger or the web, running on site. Like SAMMI.
Service 03
Facial recognition
Recognition that can run on a microcontroller smaller than your thumb, right where the camera is, even in the dark.
Service 04
Data analytics & prediction
Multidimensional data analysis and forecasting: find the patterns, flag what looks wrong, predict what comes next. Like SAMMI's 3-day price forecast.
Service 05
GPU server setup
Set up an on-site AI server so your office can run its own AI, without renting it from the cloud.
Service 06
Thesis adviser & consultation
Guidance for students and researchers working on AI, machine learning or computer vision theses, from a DLSU Golden Thesis Awardee and IEEE-published researcher.
Service 07
AI skill design for operations
Custom agent skills: written playbooks that teach AI assistants your office's own procedures, so routine work like reports, forms and checklists runs the same way every time. Skills can sync with a local AI model on site for a hybrid agentic system: sensitive work stays in-house, heavier tasks go to the cloud only with your approval.
> HOW A PROJECT STARTS
Free consultation by email or Messenger
A small pilot, scoped to one real problem
Measured results you can check
Roll out what works
> QUESTIONS
Does our data leave the office?
ANNI is designed to run on site: the AI model runs on a device at your location. If a project ever needs a cloud service, we discuss and agree on it with you first.
What does it cost?
It depends on the project, and the consultation is free. Running the AI on site means no per-message cloud fees: SAMMI costs ₱0 per message.
Which languages?
SAMMI reads Waray, Filipino and English today.
Do you work with LGUs?
Yes. SAMMI was built in Catbalogan for a regional challenge, and ANNI is designed for offices that want AI they control.
> FILE 04 · ABOUT
Personnel file
Founder
Founder & engineer · ANNI
Kahel Estrada
Michael Ralph A. Estrada
RoleAI/ML engineer
BaseManila, Philippines
In the field3 years
ResearchPublished, IEEE
Commendations
1st place, rSCENE 2026 AI Vibe Coding Challenge (Open Category), for SAMMI / Presyo Catbalogan
DLSU Golden Thesis Awardee, for a fully functional IR-supported facial recognition system
Published researcher, IEEE: "Enhancing Medical Readiness with LLMs: A Low-Resource OTC Support Bot for Deployed Units", 2025 International Conference on Mobile, Military, Maritime IT Convergence (ICMIC), Cebu. IEEE Xplore · Google Scholar
Affiliations · Philippine Coding Camp · De La Salle University · IEEE · Moody Music Company
I build AI that runs where people are: on site, in their language, under their control. ANNI started in Manila as my own assistant and has grown into a family of specialists. SAMMI was its first public deployment, built in one morning in Catbalogan and awarded 1st place at rSCENE 2026.
I'm now designing ANNI's own language model from scratch.
