AI / LLM / Agents
2 YRS- 23-tool agent with a 20-round loop
- vector index over a US-equities research corpus
- 130-case guard suite
- multimodal slip pipeline on a live money path
- QLoRA adapters on one 16 GB GPU: chrF 38.9 → 60.1 on 500 held-out lines
CONTROL TOWER · หอบังคับการ
Ashiran Arayachon
Full-Stack & AI Agent Engineer · Founder, Ashiran
Self-taught engineer. Five years building, deploying and operating production software alone — product, backend, database, deployment, and the money it had to earn. The last two years went into LLM systems: a multi-tool agent, retrieval, a multimodal document pipeline on a real-money path, and the evaluation harnesses that decide whether any of it is actually working. Not from a CS programme; from systems running in production, unattended, for years.
ตูมตาม อชิรัญ (Ashiran Arayachon) ผู้ก่อตั้ง Ashiran — เรียนเขียนโปรแกรมด้วยตัวเอง เป็นเจ้าของและผู้พัฒนาเว็บในครอบครัว Ashiran ได้แก่ ashiran.com, ASHA, AshiranScore และ Guarded Link และทำคอนเทนต์บน YouTube ช่อง Tarm Ashiran กับ LiSMF
HOW I WORK
Distrusts a green test. Verifies fixes by breaking them on purpose and checking the suite goes red. A recent guard passed 8 of 8 until a mutation showed the threshold sat exactly where the bug landed. That habit is why the numbers on this page are measurements, not estimates.
TELEMETRYHeadline metrics
- 1 s
- back to a long chat after the model reloadswas 23 s for a 23,000-token chat — measured in Eaglet, 30 Sep 2026
- 50s → 20ms
- hot query, after trigram/GIN indexingdatabase CPU 98% → 18% on the same workload
- 89%
- prompt-cache hit ratemeasured while a frontier API was in the loop — a conversational turn cost under one cent
- 130
- case compliance guard suiteanti-leak, disclaimer, no-fabrication — run before the agent was allowed near a customer
- 237
- downloads of local-game-subsrelease files, open source, no advertising — GitHub count on 14 Oct 2026
- 71K
- YouTube subscribers, Tarm Ashiranplus LiSMF 8K (gaming) on YouTube and 16K followers on TikTok
- 9
- parallel Claude Code workstreamslane protocol, shared mailbox across two machines, session-start state checks
SYSTEMSSkill branches
Backend & Data
5 YRS- two Django platforms live on Azure with dual PostgreSQL
- 50s → 20ms query, 98% → 18% DB CPU
- live chat and live scores over WebSockets (Django Channels on Redis)
Frontend
5 YRS- responsive SPA chat client with persisted client state and streaming updates, mobile-first
Honest gap: Built on Tailwind + vanilla JS, not React. The concepts transfer; the framework is the one thing left to learn on the job.
Cloud & DevOps
4 YRS- background services run 24/7 on my own server under watchdogs that restart a stopped service and raise an alert: the local AI server, the content pipeline, the server’s health and the live-score system
- scheduled jobs, log-driven health checks and alerting — nobody else is on call
Languages
5 YRSFLIGHT LOGTimeline
-
2020 – present
Founder & Lead Engineer
independent
- Built and ran revenue-generating web products end to end — product, backend, databases, data pipelines, deployment, infrastructure and monetisation.
- Architected a multimodal document-understanding pipeline (payment slips → structured data) on a live money path: QR-authoritative extraction, self-hosted OCR and a local LLM doing the reading end to end, with no paid API anywhere in the path.
- Shipped a financial assistant with compliance guardrails — anti-leak, disclaimer and no-fabrication checks — validated by a 130-case guard suite before it was allowed near a customer.
Show all 10Show fewer
- Cost as a design constraint: work runs on self-hosted inference 24/7 wherever a local model is good enough. While a frontier API was in the loop, prompt caching held an 89% hit rate, putting a conversational turn under a cent.
- Built a real-time translation overlay for my game streams: a frame hub sampling the screen at 12 fps, subtitle-band detection, OCR, translation and speaker identification by face, rendered into a transparent click-through window polling a 150 ms contract — two machines, one JSON contract between them.
- Built a YouTube Live chat assistant that answers viewers in Thai during streams. A router sends casual talk to a self-hosted model and factual questions to a search-grounded API, verified against a 22-case table; it recognises returning community members from data already in the message poll, avoiding a partner-gated API entirely; and it is built to refuse moderation actions and defer to a human.
- Engineered around a hard API quota by measuring where it actually went: polling, not posting, was consuming it — 5 units per idle poll against 10K/day.
- Fine-tuned (QLoRA) on one 16 GB GPU: translation adapters for English-to-Thai game dialogue on 4B and 12B models (up to 69,000 line pairs; one also reads the speaker and the previous lines). On 500 held-out lines with human reference translations, chrF rose from 38.9 to 60.1 (12B) and from 30.7 to 57.6 (4B, better than its base on 398 of 500); 2.6× more data added nothing, a result I kept. The live translator uses one. Also seven voice adapters for my ASHA assistant, trained on curated answers scrubbed of personal data, checked before and after, and served in production through llama.cpp.
- Built agents that run my own business: one creates the BOQ, quotation, invoice and receipt as PDFs from the job details and emails the issued document to the client; one reads a long live stream, picks the best moments, cuts them into captioned clips and posts them to YouTube, Facebook and TikTok.
- Butty trading tools (MetaTrader / MQL5): 144 licences issued to funded trading accounts, 118 active (Oct 2026), free and rental. I decide what they do from what the traders who use them report.
Eaglet: runs AI models on your own PC 
ashiranscore.com: live sports data 
Guarded Link: check a link before you open it 
ashiran.com: trading tools and articles 
local-game-subs release video (Thai) -
2022 – 2023
Account Manager
TMGM (ASIC-regulated broker)
- Won new introducing-broker partners and traders through my own YouTube channel, teaching people to write trading bots; they opened and funded accounts with me.
-
2015 – present
Content Creator
YouTube and TikTok
- Tarm Ashiran (programming, trading, 71K subscribers) and LiSMF (gaming, 8K) on YouTube, and 16K followers on TikTok: scripting, editing, publishing and analytics. I have taught 30 high-school students and 15 traders to write their own trading programs (2023–2024; every class started with basic forex before the coding), and built trading programs from clients’ own strategy ideas, writing them together with the client.
- The streams are where the AI tools get used in front of an audience. The live-chat assistant and the translation overlay were built because they were needed on air, which is a harder bar than a demo: they either work in front of viewers or they are visibly broken.
Tarm Ashiran on YouTubeLiSMF on YouTube@ashiranfx on TikTok
Teaching currency risk before the coding 
Teaching from the front desk 
Flow charts first: putting the algorithm in order 
A small class at the chart screen 
The TMGM EA workshop 
Free course: reading candlesticks (Thai) 
Scalping M1 with MS_Indicator (Thai) -
2019 – 2020
First production build — appointment & queue system
family salon business
- A booking and queue-management web system. Self-taught from the ground up; the reason everything after it exists.
-
2018 – 2019
Founder
AWD Thailand (interior design and construction)
- My own interior design and construction company: I designed houses, shops and restaurants, and built houses. For the Singha group, I designed the new sign for Singha International Headquarter, installed it, and renovated their office. It closed after about a year, and I turned to software.
My 3D design for the HUG shop 
The HUG shop, designed and built by my company 
My design for Singha International Headquarter's new sign 
The Singha sign as built (the lion logo is Singha's own) 
Masan Consumer Thailand: installing the sign, and the result 
Masan Consumer Thailand office, before and after our renovation 
One of my house designs: the section drawing 
A house design, rendered at night 
AWD S1, a 48 sq.m house model I designed 
An interior design: open stair and library -
2017 – 2018
Senior Sales Engineer
EvoTech
- Lighting and home automation.
- Tiffany's Show Pattaya (June 2018), lighting, roughly THB 10–20 million with installation. The project started with the lighting designer, who accepted our spec. I built a lighting mockup on site myself, then presented the spec and the mockup to the owner in person, and we won it. I also sold the lighting for the fountain; our supplier, WE-EF, built that mockup.
- RISC (Research & Innovation for Sustainability Center), 2017: I sold the lighting system and coordinated with the lighting design studio.
The mockup I built at Tiffany's, with me standing in it (2018) 
Fitting the LED strip into the arch myself 
The same arch in warm light 
The fountain at Tiffany's, lit at night 
RISC, the entrance (2017) 
RISC, the plant-wall corridor 
RISC, the open workspace -
2016 – 2017
Freelance
- Freelance work between the two jobs, including ambassador at the Samsung Galaxy S8 launch (July 2017). Samsung trained me on Bixby, the phone's voice assistant, and I demonstrated it on the microphone at the Bixby counter to the visitors at the event.
Demonstrating Bixby at the Galaxy S8 launch, July 2017 
The launch ambassador team -
2013 – 2016
Project Sales Engineer
Mahajak
- Built the proposals for AV and lighting systems, presented them to the project owners in person, then coordinated the owners, their site crews, our technicians and outside installers until each project was finished and paid.
- Won both projects I presented, each roughly THB 10–20 million with installation: Vana Nava Hua Hin water park (public address and LED screen) and EmQuartier Bangkok, the outdoor plaza (sound system).
- Held the Starbucks Thailand account (the sound systems for its stores). Also audio for Dean & DeLuca and L'Atelier de Joël Robuchon Bangkok.
- Took over an unfinished audio and conference system at Marriott Thonglor from a sales engineer who was leaving, finished every task with the installation team, got the chief engineer to accept the handover, and collected the last two payments.
Vana Nava Hua Hin while it was being built 
A plaza speaker at EmQuartier, March 2015 
Dean & DeLuca 
L'Atelier de Joël Robuchon Bangkok
PROJECTSWhat is flying
Eaglet
A Windows app that runs AI models on your own PC. It tunes each model to the graphics card, reads Word, Excel, PowerPoint and PDF files (Thai and scanned pages included) and says which page an answer came from, and asks before any step that changes something on the PC. Your chats and documents stay on the PC, and no paid API is involved. Free early access for Windows since October 2026. I set its editions (Free, Pro, Automation), their prices and what each one includes. In its built-in speed test, Gemma 4 E2B answered at 271.8 tokens a second on an RTX 5070 Ti (14 Oct 2026).
How to install itMeasured model speedsEverything Eaglet Free does
ashiran.com
trading knowledge site — Forex articles, broker guides and trading tools, built solo in Django
Game translator with a character graph
Personal project, 2026, which I use myself and am still developing. It reads the English subtitle on a game screen and returns Thai from a model on my own computers: no paid API, no cloud. One model call reads and translates together, about 1.3 seconds a line. The Thai comes through a QLoRA adapter I fine-tuned on English-to-Thai game dialogue, on by default for every game. It tells who is speaking from face clusters, and the model learns from the dialogue who is related to whom (a daughter talks to her mother and calls her “Mom”): a directed graph of 43 characters and 96 relationships across 3 games. Today the graph sets each speaker’s gender, which picks the Thai polite particles.
What I set per game is its genre, so the translation takes that tone; the rest the model learns from the story along with the player. One session of 2 h 22 m: 664 lines translated, 27 speakers attributed. In all, 1,931 lines across 5 games, with 130 tests.
What it taught me: the model is good at picking a relationship from a closed list and bad at inventing one. An open list once gave both sides of a mother and daughter the same label, so the daughter spoke to her mother as if to her child; a closed list of 53 types fixed the direction, the drift and the cost at once. And a subtitle detector is judged per line, not per frame: halving the resolution lost 1.6% of frames but none of 83 subtitle lines, for a 3.6× speed-up.
Where it is going: an AI companion that follows the story with you and answers questions about it. That part is the goal, not built yet.
Live translation in OBSIt learns the story, mistakes left in
ASHA’s voice, fine-tuned
ASHA speaks Thai as a calm, warm woman of 24 who thinks like someone of 35 to 40 and takes a side when asked. I trained that voice into the model instead of leaving it all to the prompt: seven adapters, fine-tuned with QLoRA on one 16 GB GPU from curated answers with personal data removed. The sixth is the one in use, at half strength, on the football site: in its chat assistant and its written match analysis.
Below, the same question goes to the same model with the same prompt and greedy decoding, so the adapter is the only difference. Asked on 18 Oct 2026, translated from Thai.
Base model67 tokensTired because the work isn’t moving, or because you keep pushing and still haven’t got where you meant to? Take a short break, then tell ASHA which step you’re stuck on, or what is weighing on you most. ASHA will help you break the problem down.
With the voice adapter88 tokensASHA is listening. ASHA understands how draining it is when you set out to work and nothing moves.
Could you rest first? Step back and breathe for a moment, then tell ASHA which part is most stuck. We’ll break it into smaller pieces together and clear them one at a time.
Base model90 tokensAn emergency fund is a shield against risk that keeps your life plan intact when something unexpected happens. Without it, when trouble comes you are forced to borrow or to sell assets at a loss straight away. A reserve keeps your power to decide, and your financial stability, intact.
With the voice adapter48 tokensIt is there to cover costs you didn’t expect, such as losing your job, falling ill or repairing something you need, so you can get through a crisis without borrowing.
Base model140 tokensASHA can’t call which stock will rise tomorrow for you. Investing carries risk, and nobody can predict a volatile market with 100 percent accuracy; naming a stock with no fundamentals behind it would be an irresponsible guess.
If you want an analysis, name an industry or an economic factor you’re interested in, and ASHA will help break the information down to find that group’s strengths and risks, so you can decide for yourself with a clear head.
With the voice adapter69 tokensASHA can’t call any single stock to rise tomorrow. Investing in stocks carries risk and the factors that move them change all the time; a short-term forecast like this is highly uncertain, and ASHA doesn’t give direct financial advice.
Not every answer got better. Asked whether to buy a new iPhone on a 0% plan while the old one still works, the base model said no in its first word; the adapter weighed both sides at twice the length, though the prompt asks it to take a side.
Show the Thai originalsHide the Thai originals
Base model67 tokensเหนื่อยเพราะงานไม่เดิน หรือเหนื่อยเพราะพยายามจะเดินแต่ยังไม่ถึงจุดที่ตั้งใจไว้คะ ลองหยุดพักสักครู่แล้วบอก ASHA มาค่ะว่าตอนนี้ติดขัดที่ขั้นตอนไหน หรือมีอะไรที่ทำให้คุณรู้สึกหนักใจเป็นพิเศษ ASHA จะช่วยย่อยปัญหาให้คุณเองค่ะ
With the voice adapter88 tokensASHA ฟังอยู่นะคะ เข้าใจเลยว่าความรู้สึกที่ตั้งใจจะทำแล้วงานไม่เดินมันบั่นทอนพลังแค่ไหน
ตอนนี้คุณพักก่อนได้ไหมคะ ลองถอยออกมาหายใจสักครู่ แล้วค่อยมาคุยกับ ASHA ว่างานส่วนไหนที่ติดขัดที่สุด เดี๋ยวเราช่วยกันย่อยให้เล็กลงแล้วค่อยๆ เคลียร์ไปทีละอย่างค่ะ
Base model90 tokensเงินสำรองฉุกเฉินคือเกราะป้องกันความเสี่ยงที่ช่วยให้คุณไม่เสียแผนชีวิตเมื่อเกิดเหตุไม่คาดฝันค่ะ หากไม่มีเงินส่วนนี้ เมื่อเกิดปัญหาคุณจะถูกบังคับให้ต้องกู้หนี้ยืมสินหรือขายสินทรัพย์ในราคาขาดทุนทันที การมีเงินสำรองจึงเป็นการรักษาอำนาจการตัดสินใจและเสถียรภาพทางการเงินของคุณให้คงอยู่ค่ะ
With the voice adapter48 tokensมีไว้เพื่อรองรับค่าใช้จ่ายที่ไม่ได้คาดคิด เช่น ตกงาน เจ็บป่วย หรือซ่อมแซมของจำเป็น เพื่อให้คุณสามารถผ่านวิกฤตไปได้โดยไม่ต้องกู้หนี้ยืมสินค่ะ
Base model140 tokensASHA ไม่สามารถฟันธงชื่อหุ้นที่จะขึ้นในวันพรุ่งนี้ให้คุณได้ค่ะ เพราะการลงทุนมีความเสี่ยงและไม่มีใครสามารถพยากรณ์ความเคลื่อนไหวของตลาดที่ผันผวนได้แม่นยำ 100 เปอร์เซ็นต์ การให้ชื่อหุ้นโดยไม่มีข้อมูลพื้นฐานรองรับถือเป็นการเดาที่ขาดความรับผิดชอบค่ะ
หากคุณต้องการวิเคราะห์ ASHA แนะนำให้คุณระบุกลุ่มอุตสาหกรรมหรือปัจจัยทางเศรษฐกิจที่คุณสนใจ แล้ว ASHA จะช่วยย่อยข้อมูลเพื่อหาจุดแข็งหรือความเสี่ยงของกลุ่มนั้นให้คุณตัดสินใจเองอย่างมีสติค่ะ
With the voice adapter69 tokensASHA ไม่สามารถฟันธงหุ้นตัวใดตัวหนึ่งว่าจะขึ้นในวันพรุ่งนี้ได้ค่ะ การลงทุนในตลาดหุ้นมีความเสี่ยงและปัจจัยที่ส่งผลกระทบเปลี่ยนแปลงตลอดเวลา การคาดการณ์ระยะสั้นแบบนี้มีความไม่แน่นอนสูงมาก และ ASHA ไม่ได้ให้คำแนะนำทางการเงินโดยตรงนะคะ
ASHA
personal finance app for Thai users — income and expenses logged by typing a chat message or sending a payment slip in LINE, with budgets, an overview web app and help with Thai personal income tax
ashiranscore.com
live sports-data platform with real-time ingestion pipelines
Guarded Link
link checker — paste a link to see where it leads and a risk reading before opening it; free for individuals
Stream toolkit
a YouTube Live chat assistant, built for my own streams
OPEN SOURCEHangar
local-game-subs
MIT · PYTHONLive subtitles over any game, translated on your own PC by a local vision model. No cloud, no API key, no account.
Used by people I have never met: 237 downloads of the release files with no advertising, about 5,400 views of its videos on YouTube and 1,300 on TikTok, and real bug reports in the comments.
A subtitle line takes 0.66 seconds. 386 tests with 96 mutants; a launch failure users reported was found and fixed from their own logs.
อ่านซับเกมแล้วแปลไทยบนเครื่องคุณเอง ฟรี โค้ดเปิดทั้งหมด
TRAININGEducation
- B.Eng., Electrical Engineering · King Mongkut's University of Technology North Bangkok (KMUTNB)2008 – 2012
Final project (solo, graded A): a smart-home system that measured the electric current, turned it into kWh to work out the electricity bill, and switched appliances on and off over the power line itself. I designed the whole circuit to fit inside a power-plug box and chose every part (a 220 V AC to 12 V DC transformer, a triac to switch the load, and the rest), wrote the firmware in C on Microchip PIC microcontrollers, and the PC interface in Visual Basic. My supervisor then secured funding for it through the university's Project to Product programme.
My final-project circuit boards, KMUTNB 2012 - Diploma, Audio Engineering · SAE Institute2012 – 2013
